Engram

Cure the collector's fallacy. Engram turns a static second brain — your Obsidian vault, your uploaded notes — into durable memory through AI-orchestrated spaced repetition.

Instead of hoarding notes you never revisit, Engram extracts concepts from them, generates MCQ flashcards via Claude, and quizzes you on a schedule tuned to how memory actually decays. Each topic appears as a living "garden" that grows when you review and fades when you neglect it.

---

Status

Active development. Phase 2 complete — AI card generation wired into the review loop end-to-end.

Phase	What	Status
Phase 1	Cloze review loop, FSRS scheduling, living dashboard	✅ Done
Phase 2	AI card generation (MCQ) with vault-sourced distractors	✅ Done
Phase 3	Metering + pricing gates	🔜 Planned

---

Stack

Layer	Tech
Backend	Java 21 · Spring Boot 3 · JdbcTemplate (no ORM)
Database	PostgreSQL 16 + pgvector (embeddings + kNN)
AI	Claude (Haiku for extraction, Sonnet for distractors) · OpenAI text-embedding-3-small
Scheduling	FSRS-6 (spaced repetition algorithm)
Frontend	Next.js 16 · React · TypeScript · Tailwind 4

---

Prerequisites

• JDK 21
• Docker Desktop (for pgvector Postgres)
• Node.js 20+ and pnpm
• Anthropic API key (sk-ant-...)
• OpenAI API key (sk-...) — used for embeddings only

---

Quick start

1. Start Postgres

docker run -d --name engram-postgres \
-e POSTGRES_DB=engram \
-e POSTGRES_USER=engram \
-e POSTGRES_PASSWORD=engram \
-p 5432:5432 \
pgvector/pgvector:pg16

2. Set API keys

export ANTHROPIC_API_KEY=sk-ant-...
export OPENAI_API_KEY=sk-...

Or add both to ~/.zshrc / ~/.bashrc to persist.

3. Start the backend

cd backend
./gradlew bootRun
# → listens on http://localhost:8081# Flyway runs migrations automatically on first boot

4. Start the frontend

cd frontend
pnpm install
pnpm dev
# → http://localhost:3000

5. Seed a concept (skip vault sync for now)

psql postgresql://engram:engram@localhost:5432/engram -c "INSERT INTO concept_candidate (concept_id, user_id, source_type, source_ref, source_content_hash, title, topic_tag, source_span, lifecycle_state)VALUES ('00000000-0000-0000-0000-000000000010', '00000000-0000-0000-0000-000000000001', 'OBSIDIAN_FOLDER', 'test.md', 'hash1', 'Spaced Repetition', 'memory', 'Spaced repetition exploits the spacing effect to improve long-term retention.', 'CANDIDATE')ON CONFLICT DO NOTHING;"

6. Open the review page

Go to http://localhost:3000/review — the app will generate an AI flashcard on first visit (takes ~5-10s while Claude writes the question and distractors), then serve it instantly on every subsequent visit.

---

Running tests

# Backend (97 tests — requires Docker for pgvector Testcontainers)cd backend && ./gradlew test# Frontend (31 tests)cd frontend && pnpm test# TypeScript checkcd frontend && pnpm exec tsc --noEmit

---

Project layout

backend/ Java/Spring service
frontend/ Next.js app
docs/
V1-spec.md Full design spec — start here for the why
progress.md Live project status
learnings.md Accumulated gotchas and conventions
CLAUDE.md Agent working rules (for Claude Code)

---

Key API endpoints

Method	Path	What
GET	/api/review/next?userId=	Next card to review (cardType: cloze|mcq)
POST	/api/review/submit	Submit a self-graded review
POST	/api/activate	Generate (or return cached) MCQ card for a concept
GET	/api/activate/{conceptId}/reveal	Fetch correct answer after user clicks Reveal
GET	/api/dashboard?userId=	Garden view with retrievability per topic

---

Design

Full spec in docs/V1-spec.md. Key decisions:

• Generate-once: MCQ cards are generated once per concept and cached. Second call returns the same card with zero LLM calls.
• Law 1 (reveal gate): The correct answer is never in the DOM before the user clicks Reveal. The /activate endpoint returns shuffled options with no correct flag; /reveal is fetched separately.
• FSRS-6: Due dates are computed from the real-valued stability interval, not whole-day rounding.
• Vault-sourced distractors: kNN neighbors from the user's own vault ground distractor generation, making wrong answers plausible to someone with shallow knowledge.