MyBrain
A personal knowledge base with semantic search.
Capture thoughts, ideas, notes, and decisions as you work. Ask Claude about them later in plain English -- MyBrain finds matches by meaning, not just keywords. Everything is stored in your own PostgreSQL database with pgvector embeddings.
Works with Claude Desktop (including CoWork mode) and the Claude Code CLI over MCP.
⚡ Easiest install: download mybrain.mcpb and drag it into Claude Desktop. See Install in Claude Desktop below. For Claude Code CLI, see Install via Claude Code CLI.
📖 New here? Start with the User Guide — a friendly tour of how MyBrain captures, scores, fades, and removes memories. Want the deep specs (schema, scoring math, conflict thresholds, the full tool surface)? See the Technical Reference.
Tool Rename (v2.0)
mybrain v2.0 implements the 8-tool atelier-brain protocol. The four legacy
tools (capture_thought, search_thoughts, browse_thoughts, brain_stats)
have been replaced by the protocol tool surface. v1 tool names no longer
exist on the server.
| Tool | What it does |
|---|
agent_capture | Store a thought with schema-enforced metadata; dedup + conflict detection + supersedes |
agent_search | Semantic search with three-axis scoring (recency + importance + relevance) |
atelier_browse | Paginated thought listing with status / type / agent / scope filters |
atelier_stats | Brain health check, counts by type / status / agent / human |
atelier_relation | Link two thoughts via a typed relation (supersedes, supports, contradicts, …) |
atelier_trace | Traverse the relation graph from a thought (backward / forward / both) |
atelier_hydrate | Ingest JSONL telemetry from a Claude Code project sessions directory |
atelier_hydrate_status | Poll completion state of a previous atelier_hydrate call |
Migration steps
-
Update tool names in your Claude Code / Cursor configs. Any custom
prompts or scripts that referenced capture_thought / search_thoughts /
browse_thoughts / brain_stats need to be updated to the new tool names
above. Old names will return "tool not found".
-
Apply the v1-to-merged migration to existing databases. v2 introduces
new columns (captured_by, origin_pipeline, origin_context,
trigger_when), new enum values (handoff, pattern, seed, plus
atelier-pipeline source agents and phases), and the thought_relations
table. The migration is purely additive and idempotent — re-running it on
an already-migrated database is a no-op.
psql $DATABASE_URL -f migrations/001-mybrain-v1-to-merged.sql
Fresh installs do not need to run this manually — lib/db.mjs's
runMigrations(pool) applies it automatically at server startup.
What You Get
Eight MCP tools Claude can call on your behalf (the 8-tool atelier-brain protocol — see Tool Rename (v2.0) above for the v1 → v2 mapping):
| Tool | What it does |
|---|
agent_capture | Save a thought with schema-enforced metadata; runs dedup + conflict detection on decision/preference types. Optional relations[] parameter creates inline edges to existing thoughts in the same transaction — collapses the common "capture + relate" pattern to one MCP call |
agent_search | Semantic search with three-axis scoring (relevance + importance + recency); refreshes recency on hits |
atelier_browse | Paginated listing filtered by status, type, agent, scope, or human |
atelier_stats | Brain health + counts by type / status / agent / human |
atelier_relation | Link two thoughts via a typed relation (supersedes, supports, contradicts, evolves_from, triggered_by, synthesized_from) |
atelier_trace | Walk the relation graph from a thought (backward / forward / both) |
atelier_hydrate | Ingest JSONL telemetry from a Claude Code project sessions directory |
atelier_hydrate_status | Poll the completion state of a previous atelier_hydrate call |
Per-call cost: SQL-only tools (atelier_browse, atelier_stats, atelier_relation, atelier_trace, atelier_hydrate_status) are free. agent_capture and agent_search each make one embedding call (free with local Ollama; ~$0.0001 with OpenRouter text-embedding-3-small). agent_capture may also make one chat call when a decision/preference lands in the conflict-candidate zone (0.7–0.9 similarity) and conflict_llm_enabled is true. atelier_hydrate cost scales with the number of telemetry files ingested.
Deeper docs:
