Memories change. Memoed finds what changed, challenges the conclusion, and opens the exact evidence behind what is current, superseded, ambiguous, or still unknown.
A search result can find an old plan. A summary can confidently repeat it. Memoed is designed for the harder question: what is true now, what changed, and why?
- Current / Superseded / Ambiguous / Unknown instead of an unsupported answer.
- Blind Challenge that searches for counterevidence independently.
- Exact sources that open the relevant audio interval or image crop.
- Local-first evidence with bounded, consented cloud synthesis.
The current build includes a bilingual, accessible Demo Data walkthrough plus local photo and audio evidence capture. Selected photos are copied into protected app storage and indexed with Apple Vision. Recorded audio is indexed on-device with SpeechAnalyzer, preserving finalized time ranges. A deterministic unified retriever applies independent evidence and character budgets, while the temporal oracle evaluates semantic assertion time, effective time, explicit corrections, conflicts, and missing context without using capture recency as truth. Empty or unavailable media is shown as an honest retryable failure instead of a false success.
Demo Data remains clearly labeled and runs without permissions or network access.
The production synthesis contract targets only gpt-5.4-mini through a server-side Responses API relay. Live cloud behavior remains disabled until its privacy, evidence integrity, and hard cost limits are verified. No API key is bundled in the iOS app.
The current walkthrough visibly discloses On-device result. The synthesis domain reserves a distinct Cloud synthesis · GPT-5.4 mini origin for a future enabled relay UI; live cloud behavior is not wired into the release flow.
- An earlier audio note says dinner is Thursday at 7:00 PM.
- A corrected invitation says Friday at 6:30 PM.
- A newer note says to follow the correction and bring a dairy-free cake.
- Memoed marks the old plan Superseded, shows the Current plan, and opens each exact source.
- Challenge independently checks whether contrary evidence should uphold, narrow, revise, or make the answer ambiguous.
Capture & Storage
→ Perception & Indexing (Apple Speech and Vision)
→ Retrieval & Temporal Oracle
→ Answer Synthesis (local fallback or bounded gpt-5.4-mini relay)
→ Citation Navigator & Blind Challenge
See Architecture, Privacy, and Validation.
Requirements: Xcode 26.6 and XcodeGen.
xcodegen generate
xcodebuild \
-project MemoedOnYourLife.xcodeproj \
-scheme MemoedOnYourLife \
-destination 'platform=iOS Simulator,name=iPhone 17,OS=26.5' \
testThe committed Xcode project is generated from project.yml. CI regenerates it and fails if the result differs.
The fail-closed relay can be verified without a credential, deployment, or provider call:
cd relay
npm ci
npm run check
npm test
npm run dry-runLive mode and both server-side spend budgets default to zero. See Relay contract.
Codex with GPT-5.6 is used as a substantive engineering partner for architecture, implementation, contract tests, debugging, device verification, and submission review. The app's runtime model contract is separately fixed to gpt-5.4-mini.
MIT