CHANGELOG.md

Changelog

All notable changes to the memex plugin. Dates in YYYY-MM-DD.

[0.17.0] — 2026-08-28

The search index leaves the vault. _index.sqlite now defaults to ~/.memex/_index.sqlite instead of living inside the vault, so a vault kept in iCloud Drive or Dropbox stops re-uploading a multi-gigabyte SQLite file every time the index is written. No schema changes, no breaking changes — an existing in-vault index keeps working where it is.

Added

• index_path config key. An optional string in ~/.memex/config.json (env override: MEMEX_INDEX_PATH) that puts the index wherever you want it — a local SSD, a scratch volume, any path outside a sync root. Leave it unset unless you need that: a set value is taken literally and skips the legacy in-vault lookup below. Its parent directory is created on first use.

• memex path --index. Prints the resolved index path; plain memex path still prints the vault path. Convenient for one-off SQL against the index: sqlite3 "$(memex path --index)" ....

Changed

• The index lives outside the vault by default. Vaults commonly sit in iCloud Drive or Dropbox, and the index is a WAL-mode SQLite file that is often several gigabytes: every write re-uploaded the whole file, sync engines left _index 2.sqlite-wal-style conflict copies behind, and Obsidian Sync was unusable on the folder. memex.paths.get_index_path() is now the single resolver. For the configured vault — the one memex_path points at — it takes the first hit of:

  1. index_path from ~/.memex/config.json, or MEMEX_INDEX_PATH
  2. <state_dir>/_index.sqlite, if it already exists
  3. <vault>/_index.sqlite, if it already exists (the legacy layout)
  4. <state_dir>/_index.sqlite — the default for fresh installs

  state_dir defaults to ~/.memex. Step 3 is what keeps existing installs working: the legacy location is still honoured, and nothing silently rebuilds. Any other vault path — tests, an ad-hoc --vault — keeps its index in-vault as before.

• Rebuild temp and backup files follow the index. _index.sqlite.tmp and _index.sqlite.bak are now written beside the index wherever it resolves, rather than landing in the vault during a rebuild.

Migration

Optional — the legacy in-vault location is still resolved, so nothing breaks if you skip this. To move an existing index out of a synced vault:

mv <vault>/_index.sqlite* ~/.memex/

The glob catches the -wal and -shm sidecars along with the database. Same volume, so the move is instant; the next command picks the index up at its new home.

[0.16.5] — 2026-08-26

Transcript frontmatter now records how a session was launched. One small feature, one deletion; no schema changes, no breaking changes.

Added

• Archive-time entrypoint stamping. Transcript conversion now reads the session JSONL's top-level entrypoint field ("cli" for interactive sessions, "sdk-cli" for SDK/fleet fan-out workers) and stamps it into the archived transcript's frontmatter. Fan-out workers running from scratch working directories mint misleading project folders; the entrypoint discriminator now survives archiving, so drift triage no longer needs to re-open the source JSONL. Stamps a scalar when the session carries one entrypoint (the normal case) and a sorted list if a stream ever mixes.

Removed

• scripts/mark_memo_saved.py. Legacy standalone script with no remaining callers — memex mark-saved (the CLI) has been the sole mark-saved path since v0.16.4.

[0.16.4] — 2026-08-26

A saved memo now stays saved. Four fixes around session state and import hygiene; no schema changes, no breaking changes.

Fixed

• memex mark-saved records the full session id, so PreCompact stops re-signaling saved sessions. The canonical state key was the 16-character session-state prefix whenever no pending signal existed — which is the normal case, since /memex:save runs before compaction. PreCompact looks up the full id, missed the prefix key, and wrote a stale "memo needed" signal for a session whose memo was already saved minutes earlier. The full id now comes from CLAUDE_CODE_SESSION_ID (when it matches the session being marked), falling back to a pending signal's recorded id, then the prefix. is_session_processed also reads legacy prefix keys (scoped to memo_generated only), so the hundreds of entries written by older versions still count as saved. The legacy scripts/mark_memo_saved.py got the same treatment.

• Empty transcripts stop re-listing as "unprocessed" forever. Sessions holding only bookkeeping lines (file-history snapshots and the like) have no conversation to convert; import now marks them skipped_empty in state and memex session discover skips them from then on. The marking runs before any --min-score filtering — a husk scores 0, so the tool's own recommended command would otherwise have dropped it before the marking path — and after --exclude, so a just-started live session you excluded can never be mis-marked. The husk check requires both a failed conversion and zero user/assistant lines (same regexes as the triage scanner); anything ambiguous stays failed, never silently skipped.

• memex session import scrubs secrets from the transcripts it writes. The import path writes the .md as the sole vault artifact (unlike SessionEnd, which archives a raw .jsonl beside it), so its write now runs through the same redaction gate as other prose writes, logging a count when anything was redacted. The SessionEnd path is deliberately unchanged.

• memex backfill memos can actually reach the Obsidian CLI. The helper lives at the vault's scripts/obsidian_cli.py, not inside the package; the packaged invocation imported a name that was never on its path, and the failure was swallowed into the regex fallback on every run. The vault scripts directory is now put on the path first. (Review caught — and reverted — a first fix that pointed at a nonexistent package module; a source-pinning test now guards against that exact wrong fix returning.)

Reliability

• memex session discover degrades gracefully when the state file is locked or unreadable instead of stalling and crashing a read-only listing.
• Pending-signal cleanup and husk marking tolerate races (concurrent mark-saved, held locks) instead of aborting mid-command.

[0.16.3] — 2026-08-05

Follow-up to 0.16.2. A latent bug removed, a flag made to do what it says, the last stray module moved into the package, and a Python floor corrected across every script that declares one.

Fixed

• The transcript converter no longer replaces its own imports with stubs. transcript_to_md imported five helpers from utils behind a two-level fallback whose "standalone mode" definitions were indented one level too far out. They therefore ran whenever the first import failed — including the case where the second one succeeded. On the import path memex session import used, Python loaded the real helpers and then immediately shadowed four of them with the minimal versions.

  Two of the four were harmless: the stub frontmatter formatter behaved identically, and the stub project-name sanitizer was never called from this module. But that sanitizer was a verbatim copy of the pre-0.16.2 function, carrying both bugs 0.16.2 had just fixed — no reserved-name table, and a length cap applied before the trim rather than after. Any future use of the name inside this module would have silently resurrected them.

  The visible half: session imports logged through the stub loggers, which print to standard error instead of the log file. Per-session import lines now land in ~/.memex/logs/ with everything else. Console output is unchanged — the import summary and any failure lines were always printed separately.

• --triage now does something when combined with --import. The import path only scored sessions when --min-score was also given, so memex session discover --triage --import accepted the flag and ignored it. Triage now runs whenever either flag asks for it, and scores appear in the import report and in --json. Filtering still happens only under --min-score, so plain --import is unchanged and pays no scoring cost.

• --exclude no longer corrupts --json. Its "excluded" notice printed to standard output ahead of the JSON document, making the result unparseable for the caller the flag exists to serve. The notice is now suppressed under --json, where an excluded session is legible by its absence.

• Scripts that declare a Python floor now declare the right one. Every hook and script carrying inline uv metadata said >=3.10, but all of them import memex, which reads its version with tomllib — a 3.11 module. A machine with 3.10 available could satisfy the declared floor and then fail at import. Because hook failures are deliberately non-blocking, that would have shown up as transcripts quietly not being archived rather than as an error. Twelve files corrected; the two that import no memex code stay at 3.10.

Changed

• transcript_to_md moved to src/memex/scripts/, leaving the usual backward-compat shim at scripts/transcript_to_md.py. It was the last module imported by a hook that still had its real implementation in scripts/, which also meant it was absent from the installed wheel — a globally installed memex reached it by inserting the vault's scripts/ directory onto the import path at call time. It is now imported like every other module and ships with the package. Both older import shapes still resolve through the shim. Several standalone tools still live only in scripts/ by design.

• --exclude filters by session id rather than comparing whole session records, removing a quadratic scan on large backlogs. Same results.

[0.16.2] — 2026-08-05

Three write-path and import-path bug fixes. No breaking changes; one additive CLI flag.

Fixed

• Sessions from hidden directories no longer land in a fragment folder. Memex recovers a session's true working directory from its transcript and validates it by re-encoding it to the ~/.claude/projects/ folder name it was found in. That validator only replaced / with -, but Claude Code's encoder also replaces dots, so /Users/you/.config/app becomes -Users-you--config-app. Every session run from a hidden directory or a worktree under .worktrees/ therefore failed validation, and project detection fell back to the lossy folder-name slug — filing the session under a fragment project instead of the real one.

  The validator now accepts either encoding. Both are needed: Claude Code has shipped two, and a single machine's project directory holds folders from both — most map every non-alphanumeric character to -, while some keep the dot verbatim. Accepting only the newer rule would have broken the folders that were already working.

• Project-name sanitizing is now idempotent, so one project cannot become two folders. The write path re-sanitizes an already-detected project name, so any input whose sanitized form changes on a second pass splits that project's content across two directories. Two such inputs existed. Reserved and unusable names resolve to _uncategorized, but running that result back through stripped the leading underscore and — since uncategorized was not itself reserved — returned it unprefixed, creating projects/uncategorized/ alongside projects/_uncategorized/. Separately, the 50-character length cap was applied after underscores were stripped, so a long name truncated onto an underscore kept a trailing _ that the next pass removed. uncategorized is now reserved and the cap re-strips, making every sanitized name a fixed point.

• memex session discover --triage --import --apply actually imports. The triage branch printed its report and returned unconditionally, so when --import was passed alongside --triage the command displayed scores and exited without importing anything — including the exact command the triage report recommends at the end of its own output. Triage now falls through to the import branch when --import is present; --triage alone still reports and exits.

Added

• --exclude SESSION_ID on session import (repeatable, accepts a unique id prefix). Skips a session at import time — intended for the session you are currently running in, whose transcript is still being written. Works with memex session discover --import and memex session import.

[0.16.1] — 2026-07-21

Fixed

• Deleting an observation now removes every row that mirrors it. observations is mirrored by three tables keyed to observation id — fts_observations (rowid), vec_observations (rowid), and observation_topics (observation_id). Two delete paths each maintained their own list of those tables and disagreed: dreamer._merge_duplicate_observations cleaned two and never observation_topics, so every duplicate merge left tag rows pointing at deleted observations.

  The damage was invisible by construction. Every JOIN-ing read path discards orphans silently, so search results were never wrong — the only symptoms were a counter reporting observations the vault could not return, and orphaned vectors consuming KNN result slots before the join dropped them. A live vault carried 515 orphaned tag rows and 187 orphaned rows in each virtual table.

  The mirror list is now a single _OBS_MIRROR_TABLES registry that both paths route through via delete_observation_ids, with a coverage test that fails at CI if a table is added to init_observation_schema without being registered. Fixing only the dreamer would have left two hand-maintained lists free to drift again.

• memex obs stats counted tag rows whose observation no longer existed. topic_observation_counts and retag_topic now JOIN observations, so their counts agree with fetch_observations_by_topic, which always joined. (memex obs topic <slug> was unaffected — it already joined.)

• delete_observations_for_doc deletes by observed id, not by doc_path. backfill obs holds a SHARED advisory lock, so a concurrent writer can insert for the same document between the id SELECT and the DELETE. A doc_path-scoped DELETE would remove that row while leaving its mirror rows behind — manufacturing the orphans this release removes. The concurrent writer's row now survives intact.

• Large deletes are chunked below SQLite's host-parameter limit. That ceiling is 32,766 since SQLite 3.32 and 999 before it, so a single-statement prune of more than 999 ids passed on a modern runtime and raised too many SQL variables on an older one.

Added

• memex obs orphans — reports index rows whose parent observation is gone, per mirror table. Read-only; --apply prunes, --json for scripts. Exits 1 when orphans exist and nothing was applied, and 2 when a mirror table could not be checked at all — a table that does not exist is omitted from tables and listed under unchecked on BOTH the human-readable and --json paths, so "not measured" never renders as "measured clean" on the surface a script consumes.

Security / correctness notes

• delete_observation_ids distinguishes a mirror table that does not exist (skip) from one that exists but cannot be read — e.g. a vec0 table on a connection that never loaded sqlite-vec, or a locked database. Both raise the same OperationalError and mean opposite things; treating the second as absence would skip that mirror's DELETE and remove the parent anyway. Existence is settled via sqlite_master; anything else aborts the delete with a message naming the fix. Mirrors are deleted before parents, so a partial failure leaves an unsearchable observation (recoverable) rather than an orphan (not detectable without this release's tooling).
• memex obs orphans --apply takes the shared writer_lock, so a concurrent index rebuild --full cannot swap the index file mid-prune and silently discard the transaction.
• The prune identifies and deletes orphans in a single statement. observations.id has no AUTOINCREMENT, so SQLite reuses freed ids; a collect-then-delete prune could destroy a legitimate tag belonging to a concurrent writer that had been granted a reused id.

[0.16.0] — 2026-07-21

Changed — BREAKING

• memex backfill obs now REQUIRES an explicit --replace or --append. Omitting both is an argparse error (exit 2) and nothing is written. The command deletes a document's existing observations in --replace mode, and for its whole life that was the silent default — so the destructive choice could be made by omission, which is how 12 observations were destroyed on 2026-07-21 by a call that printed {"stored": 5}.

  This flip is a no-op for every shipped caller. v0.15.12 deliberately shipped these flags as optional and updated commands/save.md, skills/memo-writing/SKILL.md, hooks/session-start.py, and the batch extractor to pass --replace, so the plugin cache had already rolled before the requirement landed. What breaks is an ad-hoc or external caller that omits the flag — precisely the caller this protects against.

• store_observations(...) requires a keyword-only mode argument. There is no default, and it cannot be passed positionally. The 2026-03-16 incident — store_observations called twice, the second call wiping the first — was a Python incident, so leaving a silent default in the Python API would have closed only half the historical hole. Callers passing mode="replace" or mode="append" are unaffected; callers omitting it now fail loudly with a TypeError at the call site rather than silently destroying rows.

Why this shipped as two releases

The behaviour could have been made mandatory in v0.15.12. It was staged instead because the memex CLI runs live from source while skills, commands, and hooks ship through the plugin cache — requiring a flag before the cache had rolled would have broken /memex:save and the post-compaction memo path for every user between the two events. Optional-flags-then-required is the sequence that makes a breaking change land as a no-op.

[0.15.12] — 2026-07-21

Added

• memex backfill obs now reports what it destroyed. The command has always been REPLACE-all-for-document — store_observations deletes the document's existing observations before inserting — but it reported only what it stored. Extracting 12 observations for a memo and later extracting 5 more for the same memo printed {"stored": 5, "total": 5} and left you with 5, not 17. A true statement about a call that had just deleted twelve rows, and invisible unless you diffed memex obs stats against an expected number. The output now carries replaced (rows this call destroyed), skipped_duplicate, and mode.

• --append mode, and --replace to state the default explicitly. The two are mutually exclusive. --replace remains the default so nothing breaks, but every shipped caller now passes it, which makes a future release that requires an explicit mode a no-op for them.

• A stderr warning on net row loss only (replaced > inserted). Routine re-extraction that grows a document's set stays silent by design: a warning that fires on every normal run is one operators learn to ignore, which is how the same data loss survived a "be careful at the call site" fix in March 2026 and recurred four months later.

Fixed

• Globally-deduplicated observations are no longer silently dropped. content_hash is unique across the whole table, not per document, so an observation whose text already exists under a different document — or repeated twice within one batch — was skipped without incrementing any counter. The only symptom was stored < total with no stated reason. Now counted and reported as skipped_duplicate.

• delete_observations_for_doc returns the parent DELETE's rowcount rather than the length of a preceding SELECT. backfill obs holds a SHARED advisory lock, so a concurrent writer for the same document can insert between those two statements, making the old count under-report what was actually destroyed. Callers surface this number to operators as a measured fact, and a guess presented as a measurement is the defect class v0.15.11 was spent removing.

[0.15.11] — 2026-07-21

Fixed

• memex index rebuild no longer reports a preservation result on runs where no preservation happened. An incremental rebuild printed Preserved across atomic swap: 0/0/0/0 obs/topic-tags/fts/vec rows (no prior observations) — two false claims at once, on a vault holding 15,682 observations: no atomic swap occurs on an incremental run, and the vault was not observation-less. rebuild_full set the four *_preserved keys unconditionally, rebuild_incremental never set them, and the formatter read them with stats.get(key, 0) — silently converting "this run never reported" into "this run measured zero". The line reads exactly like the silent mass-data-loss failure mode it exists to rule out, on the one command where that failure is realistic; an operator could plausibly react by restoring a backup over a healthy index, turning a false alarm into real loss.

  The reporting now models four states and distinguishes them by key presence, never by value: preservation ran and carried rows (counts); ran and the prior index was genuinely empty (counts + note); ran and threw (Preservation FAILED); or never ran — silence on incremental and first-ever full rebuilds.

• --full --no-atomic over an existing index silently destroyed observations while reporting success. That path unlinks the live index outright, with no ATTACH and no carry-over, then printed the same reassuring 0/0/0/0 … (no prior observations) line — asserting the vault had no observations on the exact run that deleted them. It now emits ⚠️ Observation preservation SKIPPED, naming how many observations were destroyed (or admitting the count could not be taken) plus recovery steps.

• All-zero preservation counts no longer claim the prior index was empty. The *_preserved counts are measured after the dangling-reference filter, so they count rows carried over, not rows the old index held. All-zero has two causes: a genuinely empty prior index, or a mass folder rename/archive done without memex obs reassign in which every observation was dropped. A new prior_observations census (taken inside the existing BEGIN IMMEDIATE + ATTACH snapshot) separates them, and the drop case now warns instead of reassuring. A partial-drop warning was added for the same reason.

• memex index status no longer renders failed counts as zeros. Each except sqlite3.OperationalError branch in get_index_status recorded 0, so a missing table or schema drift printed Observations: 0 on a vault holding thousands. Failed counts now yield None and render as unknown (query failed); an absent key renders as not reported. Relatedly, embedded_chunks no longer substitutes total_chunks as a "proxy" when sqlite-vec is unavailable — that asserted every chunk was embedded, the most reassuring possible answer, on precisely the setup where semantic search is most likely broken. It now reports unknown.

[0.15.10] — 2026-07-15

Added

• memex check now reports curator-artifact freshness. The output ends with a --- Curator artifacts --- section that reads the curator dashboard's updated: date (_meta/curator-dashboard.md) and the curator log's newest ## YYYY-MM-DD heading (_meta/curator-log.md), compares both against the newest non-archived topics/*.md updated: date, and prints a ⚠ Nd behind newest topic edit nudge when either trails by more than 7 days (CURATOR_STALE_DAYS). This catches the failure mode where the curator artifacts silently drift stale while the vault is actively tended in dev sessions that edit topics but never touch the dashboard/log. Design notes: the baseline is topic edits (not the wall clock, so a quiet vault doesn't nag); archived topics are excluded so an archive's updated: bump can't fake freshness; the log uses max() over its headings because entries aren't strictly chronological; "behind" clamps at 0. The section is hidden on vaults with no curator dashboard, and memex check --json gains an additive curator_artifacts key. No new flag — the section is intrinsic to check.

[0.15.9] — 2026-07-08

Fixed

• Bare [[project-name]] references no longer surface as false-positive ghost nodes in memex check. The crystallization/ghost-node detector resolved unresolved [[wikilinks]] against markdown filename stems + frontmatter aliases only. A project overview's file stem is _project, not the project slug, so a routine cross-project reference like [[llm-org-cognition]] or [[duality-paper]] from a sibling project's memo had nothing to resolve against and contaminated the OVERDUE/READY tiers. New _project_folder_slugs() maps every projects/<slug>/_project.md overview to its slug and merges it into the alias map on both the filesystem-fallback and Obsidian-native paths (a valid link target, just not a vote-casting source — mirroring the existing _is_archived pattern). A bare link to a not-yet-consolidated drift/fragment folder (no overview file) still correctly surfaces as a ghost. 5 regression tests added.

  (This entry was backfilled during the 0.15.10 release; the 0.15.9 code shipped and synced on 2026-07-08 but its CHANGELOG entry was missed.)

[0.15.8] — 2026-07-07

Changed

• Memo-generation subagents are now explicitly pinned to Sonnet on every fallback path. The automated Layer 2 (post-compaction) subagent already spawned with model='sonnet'; this release closes the gap on the orphan/pending-memo retry path. The SessionStart nudges (startup and resume) previously just said "Ask Claude to retry them" with no model, so a retry driven from a heavier main model could run memo generation on that model. All four nudges now instruct spawning one background model='sonnet' subagent per pending memo — memo generation is a sonnet-tier task and shouldn't burn a heavier model.
• /memex:save gained a "Model guidance" note. The inline Layer 1 flow remains primary and best-quality (the main agent writes the memo with full lived context — do not delegate to a subagent just to change models). But if you do delegate memo generation (to conserve a heavier main model or batch a backlog), pin the subagent to model='sonnet' to stay consistent with the automated fallback.

Fixed

• Docs said the Layer 2 subagent was Haiku; it has always been Sonnet. CLAUDE.md's memo-generation section is corrected to describe the background fallback as a Sonnet subagent, matching the actual hook behavior.

[0.15.7] — 2026-06-22

Added

• memex check --validate — a read-only frontmatter lint that catches the YAML-damage class: two keys glued onto one physical line (e.g. status: archivedtitle: "...", which silently keeps a doc indexed and drops its title), a dangling --- that traps the body inside the YAML block, a missing identity field (title/name), and files with no frontmatter at all. It deliberately does not enforce a status vocabulary — the vault uses a rich intentional set (evergreen, stub, developing, superseded, …), so an enum check would be pure noise; the damage class is the signal. --json for agents; exits non-zero when any issue is found (CI/launchd).

[0.15.6] — 2026-06-22

Added

• memex check --folders — a read-only audit that detects project-folder drift: cwd-fragment-shaped names (e.g. Apps-arena), duplicate/subset folders (one project's content scattered across two folders), and name≠canonical mismatches. Prints the exact obs reassign plan to consolidate (by subprefix, so it can't collide on _project.md); --json for agents; exits non-zero when high-confidence drift is found.
• project-consolidation skill — the safe SOP for merging drifted/duplicate project folders (confirm duplication → obs reassign preserving embeddings → verify the obs-count invariant), so this is self-service rather than tribal knowledge.

Fixed

• memex session import could still create Apps-* fragment folders. discover_sessions.py used the lossy Claude-dir slug parser (the same gap v0.15.5 fixed for memex sync); it now uses the canonical detect_project via the true session cwd. Canonical project detection is centralized in one place (utils) and shared by sync + import, with a tripwire that warns when a non-canonical folder would be created.

[0.15.5] — 2026-06-22

Fixed

• memex sync no longer re-fragments the vault with Apps-* folders. Auto-memory sync derived vault folder names from the lossy Claude project-dir slug (e.g. -Users-you-Documents-Apps-arena → Apps-arena), ignoring project_mappings and git remote. It now reads the true cwd from a session transcript — validated against the dir's /→- encoding so a stray path can't mis-map — and feeds the canonical detect_project(), the same identity memos use. Folder names now match the rest of the vault (arena, not Apps-arena).
• project_mappings from ~/.memex/config.json now actually applies in CLI context. get_config() returns the pydantic settings dump, but Settings had no project_mappings field, so extra="ignore" silently dropped it — detect_project()'s explicit-mapping check was a no-op outside the hook raw-json path (split-brain detection). Added the field.

Reviewed by a three-reviewer fan-out (codex + kimi + in-house). +8 tests.

[0.15.4] — 2026-06-22

Fixed

• Graph stats no longer over-count broken links from transcripts and auto-memory. extract_wikilinks — the indexer that populates the wikilinks graph table — now skips raw projects/*/transcripts/ and projects/*/auto-memory/ files as link sources and strips fenced/inline code + ANSI before scanning, matching the v0.15.3 crystallization checker. Roughly 61% of previously-reported broken links came from transcript phantoms (e.g. [[$MEMO_PATH]]). Transcripts remain valid link targets; only their vote-casting as sources is removed. Fully materializes after a memex index rebuild --full.

Changed

• New memex.scripts.wikilink_filters module is now the single source of truth for wikilink noise-filtering (strip_code_spans, is_noncurated_source), shared by the indexer, the deep-retrieval graph expansion (ask), and the crystallization checker — so the graph table and the checker can never drift again. strip_code_spans is newline-preserving, keeping per-link line_number accurate in the graph table.

[0.15.3] — 2026-06-22

Fixed

• memex check no longer launches Obsidian. The availability probe invoked the Obsidian binary to check it, which opened the app (on whatever vault was last used) when Obsidian wasn't already running — disruptive during a headless memex check. It now checks for a live Obsidian process first and falls back to the filesystem scan without launching anything. ensure_running() remains the explicit way to launch.
• Crystallization checker ignores raw transcripts as link sources. Raw conversation/terminal dumps in projects/*/transcripts/ (and auto-memory/) emitted wikilink-shaped fragments ([[$MEMO_PATH]], [[%s]], etc.) that survived code-span stripping via fenced-block edge cases and dominated the ghost-node "OVERDUE" tier. These folders are now excluded as vote-casting sources (still valid link targets), mirroring the existing status: archived skip. Real cross-project concepts still surface on curated votes; only transcript-only phantoms drop (OVERDUE 30→11 on the author's vault).

[0.15.2] — 2026-06-17

Fixed

• memex ask --depth thorough hung for minutes on a real-size index. The thorough-mode chunk vector search joined fts_content (an FTS5 table with no path index) inside the KNN query, so SQLite scanned every doc as the outer loop and re-ran the vec match per doc — O(docs × KNN). Rewrote it KNN-first (matching memex search): run the KNN on vec_chunks, push the project filter into the KNN via the v0.15.0 metadata column, then enrich titles/dates with a single batched lookup. ~1.4s now instead of hanging.

Changed

• Query embeddings are no longer pre-truncated. embed_query returns the model's native-dimension vector; each search path truncates it to whatever its vec table actually stores. This keeps vector search working during the window after you set a lower index_dimensions but before running memex index migrate-vec — previously the query/stored dimensions mismatched and search silently fell back to keyword-only.

Added

• memex index vacuum — reclaims the disk space left behind after migrate-vec drops the old larger-dimension vec tables (migrate-vec now prints a hint pointing at it). Runs VACUUM plus a WAL TRUNCATE checkpoint, so the .sqlite file actually shrinks (it's in WAL mode, where a plain VACUUM leaves the freed pages in the sidecar). Needs free disk roughly equal to the current index size.

[0.15.1] — 2026-06-17

Fixed

• Silent batch-embedding under-population. embed_content(contents=[str, str, …]) is interpreted by the google-genai SDK (both 1.x and 2.x) as the parts of a single Content, so the API returns one embedding for the whole list and the rest land as None — an N-text batch silently produced 1 vector + N−1 gaps. An older SDK auto-wrapped bare strings as separate contents, so batch embedding worked when indexes were built and regressed on a later SDK; only new embeds were affected. Fix: each text is now wrapped as its own types.Content, yielding one embedding per text (verified on the live API). Symptom this resolves: memex index embed-missing embedding ~1 item per call. Regression test: tests/test_embedding_batch_contents.py. If you ran rebuilds or backfill obs on an affected SDK, run memex index embed-missing once to fill any gaps.

[0.15.0] — 2026-06-17

Combined vector-index upgrade: Matryoshka dimensionality truncation + vec0 metadata filter-pushdown. Opt-in and fully reversible — the embedding cache keeps full-fidelity vectors, so any dimension can be regenerated without re-embedding. No behavior change unless you opt in (see index_dimensions).

Added

• index_dimensions config (embeddings.index_dimensions). Matryoshka- truncate the vectors stored in the index and used for queries to a smaller dimension (e.g. 768) while the API + cache keep the native dimensions (3072). 768d is ~4× smaller vector storage for ~0.26% retrieval-quality loss (Gemini Embedding 2 is MRL-trained). Omit it for no truncation — the default.
• vec0 metadata columns + KNN filter-pushdown. vec_chunks / vec_observations now carry doc_project / doc_type / doc_date, and search pushes --type / project / --since / --before filters inside the KNN. This fixes recall-collapse: a narrow --since no longer discards the whole semantic candidate window before filtering.
• memex index migrate-vec — migrate an existing index in place to the configured index_dimensions + metadata columns. Truncates stored vectors (no re-embed, no API calls) and populates metadata, with an atomic per-table swap. Run after setting index_dimensions.

Changed

• google-genai >= 2.0 (verified on 2.8.0). The embedding API surface (Client.models.embed_content, EmbedContentConfig) is unchanged from 1.x; the floor is raised to match where fresh installs already resolve.

Notes

• Pre-0.15.0 indexes keep working unchanged: bare vec tables (no metadata, native dim) fall back to bare KNN + post-filter automatically. To adopt the new features, set index_dimensions (optional) then run memex index migrate-vec. Do NOT run memex index rebuild before migrating — a dimension-mismatch guard warns and skips rather than re-embedding.

[0.14.3] — 2026-06-17

Ergonomics + reliability pass from an external-dependency audit (Claude Code 2.1.x, Obsidian 1.13.1, embedding stack). No breaking changes; no re-embed.

Added

• displayName in the plugin manifest ("Memex — Personal Knowledge Base"), shown in the /plugin picker (Claude Code 2.1.143+).
• outline subcommand for the Obsidian CLI wrapper. The outline() method (heading structure, --format tree|md|json) was implemented but had no argparse subcommand, so it was unreachable from the CLI — now wired. Useful for inspecting structure before condensing.
• garden-tending skill frontmatter — an argument-hint (diagnose|condense|connect|crystallize|grow|maintain) and effort: xhigh so vault-wide tending runs at full reasoning effort.

Changed

• sqlite-vec 0.1.6 → 0.1.9. Picks up proper DELETE space-reclamation (0.1.7+): incremental rebuilds and garden-tending archival churn now reclaim vector space once a chunk's worth of vectors is deleted, instead of leaving dead space in the index. No re-embed required (same 3072d float32 format); smoke-tested against the live index (hybrid + vector search verified).
• Obsidian CLI wrapper reliability. is_available(deep=True) now runs a real liveness query (vault) after the version check, so a wedged renderer (still accepts connections but returns empty for every real query) is detected and callers fall back to the SQLite graph queries instead of silently getting empty results. CLI calls are also serialized through a best-effort cross-process lock so a burst of concurrent calls (e.g. parallel garden-tending agents) can't dogpile and wedge the single renderer.

Fixed

• Atomic hook state writes. PreCompact's pending-memo signal and UserPromptSubmit's nudge state now write via temp-file + os.replace (atomic on POSIX), so a hook killed mid-write can't leave a truncated file — which PreCompact's reader would skip as corrupt JSON, orphaning the pending memo.
• Doc accuracy. The running Obsidian app is 1.13.1 (installer 1.12.4), not 1.12.5 — corrected across the rules docs, with the runtime-vs-installer dual-version scheme and a fan-out wedge-hazard note. Native Obsidian search and vault-wide tasks re-confirmed broken at 1.13.1 (async-IPC race), so FTS stays canonical. Clarified that Gemini Embedding 2 reached GA in April 2026 (distinct from the local 2026-05-07 config flip) and is the latest model.

[0.14.2] — 2026-06-16

Further precision tuning for the crystallization checker (follow-on to 0.14.1).

Fixed

• Archived files no longer cast ghost-node votes. The markdown fallback now skips status: archived files as link sources (they remain valid link targets — their filename stems and aliases stay in the resolvable set, mirroring Obsidian, since the file still exists on disk). Dead or duplicated notes (e.g. archived cwd-fragment memos) no longer inflate ref counts for concepts that should not crystallize.
• Tighter noise filtering. NOISE_REGEXES now also drops ISO-date-prefixed memo links (2026-02-16-…), explicit .md file links, @handles, and short ALL-CAPS acronyms (X, MCP, SSRN) — real topics are kebab-case. Combined with the archived-source skip, this took the author's vault from 39 to ~31 actionable ghost nodes with no loss of genuine candidates. Three regression tests added.

[0.14.1] — 2026-06-16

A bug-fix release for the Obsidian-less crystallization fallback.

Fixed

• memex check filesystem fallback no longer mis-parses code as ghost nodes. The v0.14.0 markdown fallback (scan_unresolved_via_markdown) ran the wikilink regex over raw file text including fenced code, so TOML [[section]] headers (e.g. wrangler.toml [[d1_databases]]), bash if [[ ... ]] conditionals, and ANSI terminal escapes inside code blocks were reported as actionable ghost nodes. A new _strip_code_spans() helper removes fenced/inline code and ANSI CSI sequences before scanning — mirroring how Obsidian's metadataCache parser ignores code spans. On the author's vault this dropped actionable ghost nodes from 129 to 39 (pure false-positive elimination). Two regression tests added.

[0.14.0] — 2026-06-09

A garden-tending quirk-fix release: four fixes surfaced while running a full vault-maintenance pass end-to-end (which doubles as an integration test of the plugin).

Added

• memex session reconcile-orphans [--apply] — clears stale pending-memo signals whose session already has a Layer-1 memo. A PreCompact signal persists until cleared; if /memex:save already wrote a memo, the signal is stale. A signal is "covered" when a memo exists for the same project dated within --window days (default 2). Dry-run by default; --apply deletes the covered ones, leaving genuine retries.
• memex check filesystem fallback — when Obsidian isn't running, crystallization_check now degrades to a markdown scan (resolves [[links]] against filename stems + frontmatter aliases) instead of erroring out after a 15s timeout. Makes the check usable headless/cron. All fidelity gaps over- report (a real link looks like a ghost), never the reverse — safe degraded mode.

Fixed

• **Skill !command dynamic-context injection no longer clobbers `awk` positional fields.** The harness applies slash-command argument substitution (`$1`/`$2`/`$ARGUMENTS`) to `!`command bodies before execution, so awk '{print $2}' became awk '{print }'. Replaced with cut -d: -f2. Rule: never use $<digit> inside a bang-command injection.
• Garden-tending diagnostic no longer false-flags a substantial overview that lacks memos_digested frontmatter as "never condensed" — it now reports MAINTAINED (add the frontmatter; don't re-condense).

Notes

• Documented the stale-.dist-info venv-churn cleanup as a release-SOP step.

[0.13.0] — 2026-05-25

First 0.13.x feature release. Promotes a manual SQL UPDATE pattern (used during the 2026-05-25 Apps-pi-proxy/ → pi-proxy/ folder migration that preserved 68 obs + 249 chunks across the rename, vs the morning's video-production migration that lost 43 obs to cascade-delete) to a first-class CLI command with safety guarantees the manual pattern lacked.

Added

• memex obs reassign --from-prefix X --to-prefix Y — rewrite doc_path prefix on observations + chunks in a single atomic transaction. Dry-run by default; --apply to commit; --json for scripting. Operates on the two doc_path-holding tables; sister tables (fts_observations, vec_observations, vec_chunks, observation_topics) join by rowid / observation_id, not doc_path, so a doc_path UPDATE preserves all index mirror state automatically.

• reassign_doc_path_prefix() helper in src/memex/observations.py for direct programmatic use (caller owns transaction boundaries, matches index_document / embed_chunks convention).

Safety guarantees

• Prefix-only rewrite via WHERE SUBSTR(doc_path, 1, ?) = ? + SET doc_path = ? || SUBSTR(doc_path, ? + 1). Equivalent to the manual REPLACE() pattern on clean paths but strictly safer when a folder name appears later in the path (e.g. projects/Apps-X/memos/Apps-X-backup.md), and treats SQL wildcards (%, _) inside the prefix as literal characters rather than patterns.

• Re-run footgun guard: rejects to_prefix.startswith(from_prefix) (e.g. projects/X → projects/X-old) which would compound the prefix on re-run by matching already-renamed rows.

• Invariant check on --apply: matched == updated for both tables before commit, ROLLBACK + exit 2 otherwise.

• UNIQUE collision: chunks has UNIQUE(doc_path, chunk_index). Collisions raise IntegrityError → ROLLBACK + exit 3 (data loss is worse than a failed migration).

• writer_lock wrapping prevents loss during concurrent memex index rebuild --full (matches extract.py::main and memex.dreamer convention).

• Rejects empty from_prefix and identical from_prefix == to_prefix before any SQL runs.

Tests

11 regression tests in tests/test_obs_reassign.py covering dry-run, apply, total-counts-preserved, prefix-only-substr, SQL-wildcard literals, re-run footgun, empty/identical-prefix rejection, UNIQUE collision, and no-implicit-commit. 233 total passing (was 222).

Documentation

.claude/rules/plugin-authoring.md Vault Operations section rewritten to call the CLI (with full 6-step folder-rename SOP), plus rationale for SUBSTR vs REPLACE and the UNIQUE collision exit-3 path.

Release process

Two review rounds (in-house code-reviewer + plugin-validator) before tag. Round 1 caught 2 HIGH (LIKE wildcard injection, marketplace.json version drift) + 3 MEDIUM (re-run footgun, missing writer_lock, schema-init DDL in dry-run) — all fixed in commit f4943a9. Round 2 verified the fix-set as correct, no wrong-fixes.

[0.12.2] — 2026-05-25

Wrap-up release closing four open threads from v0.12.1. Two real bug fixes (mark-saved cross-contamination, an unscrubbed write path), one helper extraction that consolidates the v0.12.1 scrub-gate pattern into a shared utility for future write-paths, and a doc-tracking cleanup.

Fixed

• memex mark-saved no longer cross-contaminates sessions. The selection heuristic (newest state file by mtime, no project filter) would mark a session in project B's signal when invoked from project A if B's state file was the most recently touched. Now prefers CLAUDE_CODE_SESSION_ID (Claude Code 2.1+ exposes this in env to any tool/CLI invoked from a session) for unambiguous selection, and falls back to newest-by-mtime only when the env var is absent. When the env var points at a session with no state file (config drift), warn to stderr and fall back to mtime rather than silently fixing the wrong session. Real-world trigger: 2026-05-25 afternoon memex mark-saved invoked from the memex cwd cleared the kimi-plugin-cc session's signal instead of memex's own.
• extract.py::append_contradictions_to_memo now scrubs before write. The contradictions-frontmatter rewrite path bypassed the PostToolUse hook (Python writes via Path.write_text, not via Claude's Write tool). Body content is normally already-scrubbed (post-v0.12.0 writes or retroactive sweep), but a v0.11.x-era memo touched here for the first time would have escaped. Now uses the new shared safe_write_text helper.

Added

• memex.scrub.safe_write_text(path, content) -> int — shared write-gate for any Python code path that writes prose to disk without going through Claude Code's Write/Edit/MultiEdit tools (the boundary the PostToolUse hook covers). Pre-scrubs content via scrub_text, writes via path.write_text, returns the redaction count. Single audit point — anywhere Path.write_text is called on user-prose content, prefer this helper.

Changed

• sync_auto_memory.py::sync_file refactored to use safe_write_text (the v0.12.1 fix is functionally identical, just shares the helper now instead of inlining the scrub call).

Tests

• 4 new tests in tests/test_mark_saved.py: env-var-driven selection picks the right session even when another session's state file is newer; mtime fallback when env var absent; warning when env var points at a missing state file; pending-memo signal cleanup.
• 4 new tests in tests/test_scrub.py::TestSafeWriteText: clean content unchanged, single-secret redaction, multi-secret redaction, OSError propagation on bad parent path.
• 222 total passing (up from 214 in v0.12.1).

[0.12.1] — 2026-05-25

Pattern-catalog expansion and one real gap closure. The v0.12.0 catalog covered the providers most likely to leak via subagent transcripts (Anthropic, OpenAI, Slack, GitHub, Google, AWS, JWT, PEM); v0.12.1 extends to four more high-confidence-shape providers, and patches the memex sync path that previously bypassed the PostToolUse hook because it writes via Path.write_text (not via the Claude Write tool).

Added

• HuggingFace tokens (hf_[A-Za-z0-9]{34,}) — covers user-access tokens and fine-grained tokens. Unique prefix; low false-positive risk.
• Stripe keys ((sk|pk|rk)_(live|test)_[A-Za-z0-9]{24,}) — covers secret, publishable, and restricted keys in both live and test mode. Uses _ (not -) so this never overlaps with the existing sk-ant-* / sk-proj-* / generic-sk patterns above it.
• Notion integration secrets (secret_[A-Za-z0-9]{43}) — 50-char total format. The 43-char alphanumeric run after secret_ is shape-distinctive enough to avoid prose collisions with phrases like secret_password.
• Sentry DSNs (https?://<32hex>(:<32hex>)?@*sentry*/<id>) — covers both the modern (public-only) and legacy (public+secret) DSN forms; matches SaaS hosts (*.ingest.sentry.io) and self-hosted Sentry installs that include sentry in the host name.

Fixed

• memex sync --apply now scrubs auto-memory content before disk. The PostToolUse hook from v0.12.0 only gates Claude's Write/Edit/MultiEdit tool invocations. memex sync writes via Path.write_text, bypassing the hook entirely — so secrets in ~/.claude/projects/<project>/memory/*.md files (which Claude's memory system writes without any scrub gate of its own) would round-trip into the vault un-redacted and become discoverable via search. sync_file() now pre-scrubs the assembled content via scrub_text(content, apply=True) before the disk write, and surfaces the redaction count in verbose output ([scrubbed: N]).

Tests

• 13 new scrubber tests: 8 pattern-detection tests covering each new provider (using runtime-concatenation fixtures so source bytes don't trip GitHub push protection on Stripe/Sentry shapes), plus 5 new prose-FP tests guarding against false matches on phrases like secret_password, hf_dataset = load_dataset(...), sk_live_demo, and the Sentry docs URL.
• 4 new sync-scrub tests covering: scrub on a single secret, scrub on multiple secrets, clean-content has no scrubbed field in the result dict, and dry-run does neither.
• 214 total passing (up from 201 baseline).

Retroactive sweep

Targeted scan (memos + topics + _meta + auto-memory + transcripts) found two pk_live_* Stripe-shape matches in a single research-starter transcript (an embedded HuggingFace publishable key from a captured API response). Publishable Stripe keys are low-risk by definition, but the scrubber is shape-based, not value-based — both occurrences were redacted to <REDACTED:stripe> for consistency with the v0.12.0 policy. No other vault content matched the new patterns.

[0.12.0] — 2026-05-25

First feature release in the 0.12.x line. Introduces a secret-scrubber CLI + library + deterministic write-time hook, triggered by an incident in which a subagent probe sequence read a local config file and surfaced three API keys into an on-disk transcript. The lesson — instruction- based controls fail in exactly the cases they're meant to catch — drove the architectural choice to make a PostToolUse hook (not subagent instructions) the primary defense.

Added

• memex scrub <path> CLI + library at src/memex/scrub.py with shims at src/memex/scripts/scrub.py and scripts/scrub.py. Detects API keys, tokens, and PEM private-key blocks via curated high-precision regex (Anthropic, OpenAI variants, generic-sk + generic-sk-vendor, GitHub PATs, Google API, AWS access keys, Slack tokens, JWT, private- key blocks). Specificity-first overlap resolution. Idempotent (<REDACTED:provider> markers don't match any pattern, so re-scrub is a no-op). Atomic write via tempfile + os.replace + fsync; preserves CRLF line endings byte-for-byte; safe on long filenames (truncates mkstemp prefix to leave NAME_MAX headroom). Exit codes: 0 clean / 1 dry-run with matches / 2 apply error.
• hooks/post-tool-use.py PostToolUse gate. Auto-scrubs every Write / Edit / MultiEdit operation targeting projects/<name>/memos/** or projects/<name>/auto-memory/**. The deterministic primary defense — doesn't depend on subagent compliance. Other paths (transcripts, topics, etc.) pass through untouched. Errors log but never block the user's write.
• New step in commands/save.md (4b — "Scrub for Secrets") and skills/memo-writing/SKILL.md (After Saving step 2) calling memex scrub --apply before observation extraction. These are the belt-and-suspenders layer; the PostToolUse hook is the primary control.
• L2 subagent prompt in hooks/session-start.py now includes explicit "do not transcribe API keys..." guidance plus a memex scrub invocation as a final guard. Path arguments are shell-quoted and the Python f-string escapes single quotes in transcript_path / project to prevent literal-corruption hazards.
• memex scrub documented in CLAUDE.md's command table.

Quality

• 54 new tests (46 for the scrubber + 8 for the PostToolUse hook). Regression tests for the overlap-algorithm specificity invariant, CRLF preservation on read+apply round-trip, long-filename mkstemp prefix, atomic-write temp cleanup on rename failure, and idempotency of <REDACTED:provider> markers.
• Three review rounds: in-house Claude code-reviewer (4 issues — 2 HIGH overlap algorithm + self-test FP, 2 MEDIUM atomic write + per-file errors — all addressed), Codex correctness audit (3 MEDIUM ship- blockers — long-filename mkstemp prefix, CRLF normalization on apply, unquoted memo-path in L2 prompt — all addressed), Claude design challenge (architectural pushback addressed by adding the PostToolUse hook), plugin-validator (PASS).

Migration

Existing memos and auto-memory files are not auto-scrubbed by the hook installation alone (the hook only fires on new writes). Backfill once:

memex scrub "$(memex path)" --apply

This runs against the full vault — memos, auto-memory, project docs, topics, and transcripts. Idempotent and safe to re-run. Transcripts are typically gitignored but still readable on disk; scrubbing them closes the local-disk exposure window.

Notes on what the scrubber does NOT catch

Line-wrapped keys (e.g. word-wrapped inside a markdown table), keys inside base64-wrapped JSON payloads, low-entropy custom-format tokens, secrets in commit messages (out of vault scope), and providers not in the catalog (HuggingFace hf_*, Stripe sk_live_* / pk_live_*, Notion secret_*, Sentry DSN URLs — slated for v0.12.1). Treat the scrubber as one layer of defense, not the only one. If you're saving a memo that intentionally discusses a secret, redact it manually rather than relying on the scrubber.

[0.11.6] — 2026-05-25

One-character bug-fix release closing a latent flaw in the v0.11.5 idempotency guard.

Fixed

• grep -Fxq flag-parsing in the topic-signal dedup. The v0.11.5 guard grep -Fxq "$SIGNAL_LINE" "$TOPIC_FILE" silently failed when $SIGNAL_LINE started with - (the bullet prefix grep parsed as a flag). The dedup never fired and duplicate signal lines accumulated on touched topics whenever the title or bullet content contained anything grep tried to interpret as an option. Fix: add -- to terminate option parsing in both grep calls in commands/save.md (the signal-line dedup AND the adjacent ## Recent signals section-existence check, hardened belt-and-suspenders even though the literal pattern doesn't start with - today).
• Skill prose pointer mirrors the requirement. skills/memo-writing/SKILL.md step 3 now mentions grep -Fxq -- explicitly so future readers transcribing the bash loop don't reintroduce the bug.

Migration

None. Existing duplicate signal lines created during the v0.11.5 window are cosmetic; the user manually deduped or left them for the next garden-tending pass. Future saves no longer hit the bug.

[0.11.5] — 2026-05-17

Bug-fix release. The /memex:save topic-stamping step in commands/save.md was not idempotent: when a memo's frontmatter listed multiple slugs that all redirect to the same canonical topic (e.g. claude-code-plugins, plugin-architecture, plugin-development all → Claude-Code-Plugins), the bash loop appended the same signal line once per input slug. One save could leave two- or three-times duplicates on the canonical topic file. The same pattern hit multi-agent-code-review → multi-agent-review. No data loss — just visual noise in ## Recent signals sections.

Fixed

• Idempotent topic stamping. The bash loop in commands/save.md step 6 now wraps the append in grep -Fxq "$SIGNAL_LINE" "$TOPIC_FILE" so the same line is never written twice to the same topic file. -F treats [[, |, ]]), . as literal (no regex surprises); -x requires whole-line match (no substring false positives). Also covers the case where /memex:save re-runs on an existing memo across sessions.
• Skill instruction mirrors the requirement. skills/memo-writing/SKILL.md step 3 now explicitly states the dedup requirement so out-of-band invocations (skill triggered without going through the slash command) inherit the same behavior.

Migration

None. Existing duplicate signal lines in topic files are cosmetic; clean them by hand when convenient or let the next garden-tending pass handle them. Future saves will no longer create new duplicates.

[0.11.4] — 2026-05-13

Hotfix for v0.11.3 across four rounds of multi-reviewer audit (3+codex on the v0.11.3 ship, then 4 on the fix-set, then 1 narrow gate on the final delta — 22 + 10 findings total). The audit caught two HIGH-severity data-loss paths in v0.11.3 and several class-of-bug repeats; this release hardens all of them. No schema changes — drop-in upgrade.

Fixed

• Partial observation preservation no longer ships a half-baked index (HIGH). v0.11.3's rebuild_full --atomic caught sqlite3.OperationalError during ATTACH-old preservation and only logged to stderr, then the finally block committed and the atomic swap proceeded. A disk-full partway through the four INSERT…SELECT statements would have produced a "successful" rebuild with broken observation FTS/vector search and no error signal — strictly worse than the May 7 wipe (which at least failed loudly). Preservation now runs inside SAVEPOINT obs_preserve; any exception triggers ROLLBACK, sets stats["preservation_error"], raises RuntimeError with a stable prefix, and the CLI exits 4 with a clear message instead of falling through to the swap.
• Fresh-install slash commands no longer break out of the box (HIGH). README/SETUP Quick Start used to install the plugin before the memex CLI, so the first /memex:status after install failed with memex: command not found. Quick Start is reordered: uv tool install is now Step 1, plugin install is Step 2.
• redirect_to: resolver handles cross-namespace targets (P0). The bash resolver in commands/save.md hardcoded $VAULT/topics/$slug.md, so real-world redirects like topics/bloom.md → projects/clawd-world/_project.md silently dropped signals. Resolver extracted to a Python module + memex topic resolve CLI subcommand. Bare slugs resolve under topics/; targets containing / resolve as vault-relative paths (with auto .md suffix). Path traversal (../../etc/passwd) is blocked via Path.resolve() + containment check.
• Cycle detection in redirect chains (P0). The 5-hop limit is now a fallback, not the primary guard. The resolver tracks a visited-set and reports cycles as WARN: redirect cycle detected: A -> B -> A with the actual chain — not the previous generic "exceeded 5 hops" message.
• CHANGELOG migration note for v0.11.3 was factually inverted. v0.11.3's text said terminal-archive (status: archived without redirect_to:) would "silently land on the archived stub". Actual behavior is "skip with stderr warning". The note is rewritten to describe both archive shapes correctly.
• batch_extract_observations.py exit code reflects partial failure. Previously exited 0 if any single memo succeeded — a run of 1-ok + 99-store-failed exited 0, hiding data loss from automation. Now exits 0 only when all results are ok-or-skipped, 1 when no success, 2 when partial failure (matches the v0.11.0 backfill obs convention).
• commands/status.md no longer hardcodes ~/.memex/pending-memos. The path resolves through state_dir config, which the hardcoded pipeline ignored. Replaced with memex context lookup.
• init_observation_schema no longer commits internally. Restores the v0.11.1 "callers own transactions" convention. All call sites audited; each already commits downstream.
• config.json.example cleanup: prior fix mistakenly claimed project_mappings was an unused phantom field. It's actively read by detect_project() in src/memex/scripts/utils.py as priority-1 project name override. Example block restored with accurate comment.

Changed

• Observation preservation refactored to registry + helper. Module constant _OBS_PRESERVATION_TABLES lists every preserved table. Test test_preservation_registry_covers_init_schema enforces that any new table added to init_observation_schema either appears in the registry or in the documented FTS/vec special-case branches — converts the v0.11.3 fts_observations bug class from runtime-symptom to CI-time discovery.
• memex topic resolve <slug> is a first-class CLI subcommand (registered under the new topic Typer group). Used internally by /memex:save and available standalone for testing or scripting redirect-aware tools.

Added

• Advisory file lock at ~/.memex/locks/full-rebuild.lock prevents memex backfill obs and memex.dreamer from racing against an in-progress memex index rebuild --full. New helper memex.db_utils.writer_lock() (context manager, LOCK_SH | LOCK_NB) wraps observation-write call sites in extract.py and dreamer.py. --full rebuild acquires LOCK_EX | LOCK_NB; contention exits 3 with a clear message. Also: BEGIN IMMEDIATE on the old DB before ATTACH belt-and-suspenders against stragglers that haven't yet hit the lock check.
• tests/test_topic_resolve.py — 24 tests covering the redirect resolver: slug/path resolution, cross-namespace targets, cycle detection, quoted/comment/whitespace edge cases, path traversal blocked, binary-target graceful handling, empty frontmatter fall-through, auto-.md suffix on path-form targets.
• tests/test_batch_extract_observations.py — 13 tests for parse_json_array (fenced JSON, prose preamble, trailing prose, nested arrays, empty input, whitespace-only).
• Additional tests in tests/test_index_rebuild.py: registry coverage, vec_observations preservation across atomic swap, observation_topics broken-ref filter, preservation-failure aborts swap, CLI exits 4 on preservation failure, CLI re-raises unrelated RuntimeErrors.
• FTS5 rowid invariant documented in code at both the INSERT INTO fts_observations site (extract.py) and the CREATE VIRTUAL TABLE site (observations.py). Future maintainers who change either side know the preservation path depends on the equality fts_observations.rowid == observations.id.
• memex index status listed in CLAUDE.md CLI commands table (was documented in prose only).

Migration notes

• Existing installs: no action required. The fixes are internal hardening + UX tightening. A future memex index rebuild --full will exercise the new SAVEPOINT-protected preservation path.
• If running automation that invokes memex backfill obs or memex.dreamer while a rebuild may be in progress: the new lock causes the backfill/dreamer to exit 3 on contention rather than silently lose writes. Wrap automation in retry-with-backoff if you need transparent recovery.

---

[0.11.3] — 2026-05-13

Three fixes addressing curator-tending findings: redirect-aware signal routing, fast memex status on large indexes, and observation preservation across atomic full rebuilds. No schema changes — drop-in upgrade.

Fixed

• memex status no longer hangs at 100% CPU on large indexes. count_embedding_gaps was running LEFT JOIN chunks × vec_chunks against the sqlite-vec virtual table — the join can't use an index, so it probed per-row (~107s on 136K chunks). Replaced with a COUNT(chunks) - COUNT(vec_chunks) delta — ~2000× faster (107s → 53ms). The per-doc DISTINCT query only runs when a gap actually exists (rare
  • actionable). Observations LEFT JOIN stays (low cardinality).
• rebuild_full --atomic preserves observations across the swap. Background: observations are extracted by sonnet subagents (memex backfill obs --stdin), not derived from documents at index time. The May 7, 2026 incident on the maintainer's vault wiped ~3265 observations because the atomic swap installed a fresh empty DB. Recovery cost ~$10-20 in subagent time. Now rebuild_full runs ATTACH DATABASE old → INSERT...SELECT for observations + observation_topics + vec_observations, filtered to doc_paths still present in the new fts_content table. Dangling refs (memos deleted between rebuilds) are intentionally dropped. Defensive: handles pre-0.11 schemas without obs tables, commits before DETACH so the unlock succeeds.

Changed

• /memex:save now follows redirect_to: chains when appending "Recent signals" to wikilinked topics. When archiving a topic whose content was absorbed into a canonical replacement, set redirect_to: <target-slug> in frontmatter (alongside status: archived). Signals from new memos will route to the target instead of accumulating on the dead-end stub. The resolver walks up to 5 hops, then bails with a warning. Topics archived without redirect_to: (typical for project-target archives where content belongs in a _project.md) get a WARN: ... skipping signal message — by design, not a bug.
• skills/memo-writing/SKILL.md step 3 updated to match: "Resolve redirect_to: in frontmatter first so archived topics route to their canonical replacement."

Added

• scripts/batch_extract_observations.py — async dispatcher that fans up to 5 concurrent claude --print --model sonnet subprocesses across a list of memos to populate the observations table. Useful for users who imported existing memos before observations existed, or who need to refresh observations en masse. Idempotent: skips memos that already have observations. Logs per-memo JSON results to ~/.memex/logs/batch-obs-extraction.jsonl. Rate ~0.12 memos/s at 5-way concurrency.
• 4 new tests in tests/test_index_rebuild.py (98 tests total, 97 pass + 1 skipped):
  • count_embedding_gaps fast path: asserts no LEFT JOIN against vec_chunks when total == embedded.
  • count_embedding_gaps fallback: per-doc query DOES run when gap > 0.
  • rebuild_full preserves observations across atomic swap.
  • rebuild_full drops obs for deleted memos (filter correctness).

Migration notes

• Existing installs: nothing required. The redirect_to: convention is opt-in, but the two archive shapes now behave differently:
  • Redirected archive (status: archived + redirect_to: <slug>) — the resolver follows the chain (up to 5 hops) and the signal lands on the canonical replacement.
  • Terminal archive (status: archived, no redirect_to:) — the resolver emits a stderr warning (WARN: archived with no redirect_to — skipping signal) and the signal is dropped, not silently landed on the archived stub. To preserve signals for archives whose content moved elsewhere, add redirect_to: to the archive's frontmatter as you encounter them during garden-tending.
• memex status speedup is automatic.
• Observation preservation is automatic on the next --full rebuild. No more "I just rebuilt and now memex obs stats is empty" surprises.

---

[0.11.2] — 2026-05-07

Throughput pass on the embedding pipeline. No schema changes, no breaking API changes — drop-in upgrade. After installing, no rebuild required; your next memex index rebuild --incremental (and the nightly job) will just be much faster.

Changed

• Default Gemini model flipped from gemini-embedding-2-preview to the GA gemini-embedding-2. Both names produce 3072-dim unit-norm vectors and are interchangeable — the GA name is just the documented stable identifier. Existing embeddings keep working; new embeddings come from GA. If your ~/.memex/config.json pinned the -preview name, it still works; switching is optional.

Added

• Async/concurrent embedding dispatch. Sub-batches now run in parallel under asyncio.Semaphore(GEMINI_CONCURRENCY=5) instead of sequentially with a 1-second inter-batch sleep. Gemini Tier 2 paid is TPM-bound at 5M tokens/min; with 8K-token batches the realistic ceiling is ~625 RPM, so 5 concurrent ≈ 48% utilization with headroom. Expect order-of-magnitude speedup on --full rebuilds and noticeably faster nightly --incremental runs.
• ±20% jitter on retry backoff. GEMINI_BACKOFF_SCHEDULE is now randomized per attempt to prevent thundering-herd retries when multiple concurrent batches all hit 429 at the same instant.
• Running-event-loop fallback for the sync embed_texts API. If invoked from inside an async context (the way scripts/mcp_server.py tool handlers would call it), embed_texts punts to a worker thread with its own loop instead of raising RuntimeError: asyncio.run() cannot be called from a running event loop.
• threading.Lock around _get_client. The env-var stash dance that suppresses the SDK's "both keys set" warning is now serialized across threads.
• Two new regression scenarios in scripts/verify_embedding_retry.py: concurrent_multi_batch (force 3 sub-batches; assert order preservation + zero inter-batch sleeps) and inside_running_loop (the MCP-server regression we caught and fixed).

Behavior change worth knowing

• Partial failure is more granular. Pre-0.11.2, a single bad sub-batch would None-pad itself and every subsequent batch. In 0.11.2, parallel batches launch all at once; one batch's failure no longer kills its siblings — only the failing batch's slots become None, successful batches' vectors survive in the result list. PartialEmbeddingFailure is still raised so callers know something failed; positional alignment to input texts is preserved.

Internal

• GeminiProvider.embed_texts now wraps _aembed_texts (async). The public API stays sync.
• The async path uses await asyncio.to_thread(client.models.embed_content, ...) rather than client.aio.models.embed_content directly, because the SDK's aio HTTP client binds to whichever loop first touched it and goes stale across our running-loop fallback's loop boundaries.

---

[0.11.1] — 2026-04-21

Hardening pass on top of 0.11.0. Critical transcript-chunking fix.

Fixed

• Critical: transcript chunking slice bug. chunk_transcript_turns used parts[1::2] against a non-capturing regex. re.split with a non-capturing pattern returns [pre, body_1, ..., body_N] — no interleaved delimiters. The old slice silently dropped every other turn and paired surviving bodies with the wrong headers. Net impact: every multi-turn transcript indexed before this fix had ~half its turns missing from the semantic search index, with mismatched headers on the rest. FTS5 was unaffected (tokenized raw content independently), which is why nothing surfaced as an error. Existing installs must run memex index rebuild --full after upgrading to re-chunk every transcript correctly.
• Rollback test in test_index_rebuild.py was vacuous — monkeypatched the wrong function. Now monkeypatches index_document (which runs after FTS is written) and asserts the FTS row stays at v1 content.
• enable_load_extension(True) no longer leaves the connection in a privileged state if sqlite-vec fails to load (try/finally).

Added

• Unit-norm invariant on embedding provider output. Both GeminiProvider and LMStudioProvider call _assert_unit_norm at the end of embed_texts. Samples head+mid+tail of each batch, raises ValueError if ||v||² drifts more than ±0.02. sqlite-vec uses L2 distance by default; only monotonic with cosine when inputs are normalized. New providers must call this helper.
• memex.db_utils module. connect_index(path) applies journal_mode=WAL, synchronous=NORMAL, busy_timeout=10s so writers don't collide with parallel memex backfill obs --stdin. load_vec_extension(conn) loads sqlite-vec and clears the privileged state in finally. Use this everywhere instead of bare sqlite3.connect(path).
• Per-doc SAVEPOINT pattern in rebuild loops. rebuild_full and rebuild_incremental wrap each doc in SAVEPOINT doc → work → RELEASE SAVEPOINT doc, with ROLLBACK TO SAVEPOINT doc on exception. Split into _rollback_savepoint_or_die (propagates any ROLLBACK error) and _release_savepoint_if_exists (tolerant only of "no such savepoint" substring). Prevents partial commits on mid-doc exceptions while ensuring real SAVEPOINT bugs still surface.
• Nightly rebuild support. scripts/nightly-rebuild.sh handles incremental rebuild + embed-missing retry, sources ~/.secrets, survives malformed secrets with a clear error. See SETUP.md for a launchd/cron snippet.
• Hermetic test suite shipped to public repo for the first time (11 files, 93 passed + 1 skipped on a fresh install).

Changed

• index_document and embed_chunks no longer call conn.commit() — transaction boundaries belong to the caller. Single-file CLI callers commit explicitly; batch callers use the SAVEPOINT pattern above.

[0.11.0] — 2026-04-20

Added

• memex index embed-missing command. Idempotent retry path for when rebuild completes but the embed step failed (expired API key, rate limit). Finds chunks/observations in the index but missing from vec_chunks/vec_observations via LEFT JOIN, re-embeds through the pipeline. Exits non-zero if any remain unembedded.
• embedding_gaps surfaced in format_rebuild_stats + format_status.
• Dual Gemini/Google API key stderr-noise handling (env stash/restore in _get_client).
• PartialEmbeddingFailure typed exception. After 4 attempts (initial + 3 retries with 10s/30s/90s backoff) on 429/500/503, GeminiProvider.embed_texts raises this with partial results attached. EmbeddingPipeline.embed_text (single-text wrapper) catches it and degrades to None for query-path resilience; embed_chunks caches partial success and logs failures to stderr.

Removed

• pending_embeddings.jsonl queue. enqueue_embedding_job / dequeue_embedding_jobs / get_embedding_queue_count were dead code (zero call sites) and duplicated embed-missing. Removed.
• memex context --full and --compact. The rich context injection surface backed by src/memex/context.py was removed by design. recall skill DEEP mode replaces it. The memex context CLI command is preserved but trimmed to project-detection + pending memo status.

[0.9.0 – 0.10.0] — 2026-04 (consolidated)

Added

• Observation-to-topic clustering. observation_topics junction table links observations to topic slugs (many-to-many). Enables bounded complete retrieval: memex obs topic <slug>, memex obs stats, memex obs retag <old> <new>, memex obs untagged.
• memex backfill topic-tags propagates memo-frontmatter topics to observation_topics.
• Gemini Embedding 2 support (3072d). Dimension migration auto-drops vec_chunks on provider switch (1024d LM Studio ↔ 3072d Gemini). Preview model gemini-embedding-2-preview does NOT support task_type — _build_embed_config strips it.
• Scoped observation search and similarity detection (memex similarity).

[0.7.0 – 0.8.0] — 2026-04 (consolidated)

Added

• Typer-based unified CLI. Subcommand groups: index (rebuild, status, embed-missing), obs (topic, stats, retag, untagged), backfill (obs, tokens, memos, topic-tags), session (discover, import), graph (backlinks, orphans, tags, stats).
• memex read, memex path, memex check, memex context.

Changed

• Primary interface shifted to CLI + skills. scripts/mcp_server.py still ships as an optional MCP integration, but the recommended path is now the memex CLI (shell access) plus intent-based skills (recall, garden-tending, memo-writing, curator-practice) for in-session use. New features land in the CLI first.

Removed

• scripts/batch_generate_memos.py (the only import anthropic in the codebase, contradicted the "no external API calls" architecture).

Breaking Changes Summary (0.6.0 → 0.11.1)

If you're upgrading from 0.6.0, expect these user-visible changes:

1. Slash commands removed: /ask, /backfill, /load, /maintain, /merge, /retry, /search, /synthesize, /timeline. Use the equivalent CLI (memex ask, memex search, memex timeline, memex backfill) or skill-based flows (garden-tending handles the former synthesize and merge behavior). Kept slash commands: /memex:save, /memex:status, /memex:open.

2. memex context --full and --compact removed. Rich context injection is no longer in the CLI. The recall skill (DEEP mode) covers cross-session synthesis.

3. ask-memex skill removed (absorbed into recall).

4. Transcript re-chunking required: after upgrading, run memex index rebuild --full once to fix the chunking slice bug (see 0.11.1 "Fixed").

5. Embedding provider change: if you were on LM Studio (1024d) and switch to Gemini (3072d), the migration path auto-drops the vec_chunks table on first rebuild. Same in reverse.