Skip to content

PRODUCT_BRIEF.md

Latest commit

 

History

History
63 lines (57 loc) · 3.84 KB

PRODUCT_BRIEF.md

File metadata and controls

63 lines (57 loc) · 3.84 KB
Raw

Product Brief

What Alice Is

Alice is the continuity layer for AI agents: a local-first memory service that lets agents resume interrupted work, track open loops, recall decisions with provenance, and improve when corrected. Agents connect over MCP, HTTP API (per-agent API keys), or CLI.

Customer

Agent developers — people building or operating AI agents who need durable, explainable memory across sessions. A human-facing knowledge product may come later, built on the same surface; it is not the current target.

What Makes It Different

  • Typed continuity objects: decisions, open loops, resumption briefs — not just extracted facts.
  • Explainable provenance: source-backed memories trace to source evidence where present, while reviews and corrections preserve their own audit chain. Explicit commits may legitimately have no source reference.
  • Review-governed writes: agent commits resolve to commit, confirm, review, or reject through policy; the review console is the trust boundary.
  • Local-first: your data stays on your machine; models and embeddings are pluggable via OpenAI-compatible endpoints.

Current Posture

  • v0.14.0 is the latest published release and immutable baseline (Phase 5 enterprise track: real-host single-tenant deployment contract and least-privilege operations). v0.13.1 is the prior published release. It carries the Phase 4 core-roadmap work (replicated benchmark, SQLite scale, reference integrations). Its PyPI wheel and source distribution have Trusted Publishing provenance, with exact digests in docs/release/v0.14.0-checksums.txt.
  • v0.11.0 shipped the Phase 1 periphery cut. It removes the Telegram, hosted/control-plane, public chat/response, chief-of-staff, and model-pack periphery so the default runtime matches the product described here. The low-level response job machinery remains internal to retained provider invocation.
  • The product baseline includes hybrid retrieval, the eleven-tool core MCP surface, the Memory Operations Protocol, per-agent API keys, Context API v2, temporal graph memory/entity resolution, and six live eval suites.
  • The 79.4% LongMemEval_s (397/500) figure is one historical run from 2026-07-07 with per-question evidence. It is not a repeated-run estimate or a measurement of the current release.
  • The fourth-audit remediation shipped in v0.10.3 after exact-SHA gates and independent review passed.
  • Alice is public-alpha, pre-1.0, single-user, and self-hosted. alice-memory is published on PyPI; uvx alice-memory mcp serves the core tools against a local SQLite file with no Docker or Postgres.
  • v0.12.0 shipped the Phase 3 structural refactor with Structure only. Zero behavior change. It relocates oversized HTTP, store, contract, MCP, and CLI modules behind stable facades and entrypoints without changing the local-first product boundary.
  • Phase 4 shipped as v0.13.1. Phase 5 (enterprise track) has not begun.

Non-Goals (Now)

  • Hosted service, multi-tenant control plane, SLA, or managed cloud.
  • Telegram or other channel transports.
  • Managed OAuth consent/account-linking flows or automatic account syncing. Temporary manual-token Gmail and Calendar compatibility is unmounted by default behind ALICE_LEGACY_SURFACES=1.
  • A public bundled chat/response product, chief-of-staff product, or model-pack catalog.
  • Automatic memory capture from arbitrary conversation.
  • Marketplace, channels, or vertical-agent products.
  • OCR or transcription execution; Alice ingests text extracted elsewhere.

Success Criteria For The Overhaul

  • An agent developer can go from clone to a working MCP-connected memory in one quickstart path.
  • Search quality is measurably better with embeddings configured, and honestly labeled when degraded.
  • The MCP surface is small enough to learn in one sitting.
  • Every claim in the docs matches shipped behavior.