ai-memory

Auto-generated from project docs

North Star

Booklet Status

BookletStatusBranchAppetiteDeadlineNotes
B0COMPLETEN/A1hr2026-03-30Inherited. MEMORY.md cleanup, dead systems removed
B1COMPLETEN/A1 day2026-03-30Inherited. Mem0 + pgvector + MCP + 31 memories
B2COMPLETEN/A1.5hr2026-04-0728→4 feedback files, 9 dead files killed (83→59), MEMORY.md 131→96, enforcement hook, Mem0 wired into session-manager.sh, mem0-cli.sh
B3COMPLETE45min2026-04-07Min Min verify-exec.ts wired (recall+remember), HAA run-with-mem0.sh (pre-fetch+post-store), claw_memory migrated (32→39), cross-agent test passed
B4COMPLETE20min2026-04-07memory-audit.sh, cron Monday 8UTC, TTL warning, mem0-cli list/delete
B5CONFIRMED-WORKING4hr2026-04-12v2: MCP memory server + Postgres schema + migration + auto-bootstrap. 102 memories, Grade A. All 6 MCP tools native-callable from Claude Code (verified 2026-04-13). INCIDENT-049 resolved.

Recent Decisions

DateDecisionRationale
---------------------------
2026-04-12v2 rescope: database-backed MCP serverINCIDENT-044 + RESEARCH-224. File-based index fails repeatedly. Every mature system uses DB as source of truth.
2026-04-12Custom pgvector over keeping Mem0Mem0 summarizes on write (lossy). Need raw storage + hierarchical schema. Same pgvector infra.
2026-04-12Store raw, retrieve smart (Mem-Palace pattern)Never summarize at ingestion. Structural filtering + hybrid search on read. 34% retrieval boost.
2026-04-07Rescoped from mem0-memory-layer to ai-memoryBroader scope: unified memory, not just Mem0 install. RESEARCH-206 gap analysis.
2026-04-07Two layers, not fiveCLAUDE.md/MEMORY.md for bootstrap + Mem0 for shared knowledge. Everything else is noise.
2026-04-07Enforcement via hooks, not disciplineINCIDENT-024 proved agent discipline fails. Hooks are the only reliable enforcement.
2026-03-30Mem0 over custom build48K stars, Apache 2.0, native pgvector. Why reinvent?
2026-03-30HAA read-only initiallyAutonomous agent writing bad memories propagates to all agents. Validate first.
2026-03-30Gemini embeddings over localFree tier covers volume. Saves RAM vs Ollama.

Source: /root/projects/ai-memory/