Status: Archived — Lessons learned have been upstreamed into Animus. See
animus/memoryfor the spiritual successor.
memboot was an experiment in local-first, project-aware memory for AI coding assistants.
It explored ideas that were under-served in 2025–2026:
- AST-aware chunking — indexing Python functions, classes, and methods as semantic units
- Episodic memory primitives — decisions, notes, observations, and patterns with bitemporal tracking (
valid_from/recorded_at) - Security-first indexing — SSRF guards, credential skip during chunking, SQLCipher encryption
- Zero-infrastructure vector search — SQLite + TF-IDF with no Docker or external DB required
- MCP-native integration — stdio JSON-RPC server for Claude Code, Cursor, Windsurf, Zed, and other AI IDEs
memboot succeeded as a learning vehicle but failed as a standalone product:
- Its vector search was inferior — SQLite + TF-IDF lost to Chroma on latency, scale, and retrieval quality in every dimension.
- Its moat was thin — AST chunking and regex credential skip are moderate-effort features for a funded competitor.
- The right home for these ideas is Animus — a sovereign AI operating environment, not a standalone CLI tool.
The valuable parts of memboot have been extracted and upstreamed into Animus:
| memboot Component | Animus Destination |
|---|---|
chunker.py (AST parsing) |
animus.ingest.code |
models.py (Memory types) |
animus.memory.types |
security.py (SSRF, encryption) |
animus.security.guards |
mcp_server.py |
animus.mcp.memory_server |
query.py (search logic) |
animus.memory.retrieval (backed by Chroma) |
CLI workflow (init, remember, query) |
animus CLI subcommands |
- AI agents need project-aware memory — codebases are graphs, not bags of files. AST chunking matters.
- Security is not optional — credential skip and SSRF guards should be defaults, not afterthoughts.
- Bitemporal tracking is underrated —
valid_fromvsrecorded_atlets you answer "what did we believe at the time?" - MCP is the right integration layer — stdio JSON-RPC is simpler and more robust than HTTP APIs for IDE plugins.
- Don't build your own vector DB — Chroma, Qdrant, and pgvector exist. Use them.
Apache 2.0
For the future of local-first AI memory, see Animus.