BrainMemory-MCP

A Model Context Protocol (MCP) server that gives AI/LLM agents a durable brain memory — the ability to store, recall, search, connect, summarize, and forget information across sessions through standardized MCP tool calls.

Since v0.4.0 memory is modelled internally as a small knowledge graph:

• memories are the graph nodes (content, category, tags, importance),
• connections are directed links between memories (e.g. related_to, caused_by, part_of),
• details are extra facts attached to a single memory.

This makes recall precise — instead of only matching words, the server can walk the connections you build (multi-hop recall) and explain how two memories relate (shortest path). The tool vocabulary stays "memory"-oriented (no "entity" wording), and it is still just SQLite under the hood — zero extra dependencies.

The server runs in two modes:

• stdio (default) — the server is launched as a subprocess by an MCP client (e.g. via uvx brainmemory-mcp).
• web — MCP over HTTP + Server-Sent Events (SSE) with --web, so remote MCP-capable clients (Claude, IDE agents, etc.) can connect over the network.

Memory is persisted locally under ~/.brainmemory-mcp (a SQLite database).

Cognitive Tools (12)

Since v0.9.0 the tool surface is consolidated: every operation takes a list, so acting on one memory or fifty is the same call (a single item is just a list of one). Detail and link writes are unified into one mixed-operation batch tool per entity. The result is full CRUD over all three entities with just 12 tools.

Every list-taking tool processes items independently and reports a per-item status — one bad item never aborts the batch.

Mixed-operation batches

edit_details — each item's op selects the operation:

{"items": [
  {"op": "add",    "memory_id": "<id>", "content": "config lives in /etc/nginx"},
  {"op": "update", "detail_id": "<id>", "content": "corrected fact"},
  {"op": "delete", "detail_id": "<id>"}
]}

edit_links — connect/disconnect memories, mixed in one call:

{"items": [
  {"op": "link",   "from_id": "<a>", "to_id": "<b>", "relation": "depends_on", "weight": 0.9},
  {"op": "link",   "from_id": "<a>", "to_id": "<c>"},
  {"op": "unlink", "from_id": "<a>", "to_id": "<d>"}
]}

Re-linking the same from/to/relation updates the weight (upsert). Detail ids are returned by the add op and by recall_memories(include_details=true).

Two read-only resources are exposed as JSON: brainmemory://stats (the summary) and brainmemory://graph (the nodes + links map).

Migrating from v0.8.0 or earlier: the singular tools (store_memory, recall_memory, update_memory, forget_memory, add_detail, link_memories, unlink_memories) and the v0.8.0 bulk names (store_memories kept its name; add_details, link_memories_bulk, unlink_memories_bulk were folded into edit_details / edit_links) are replaced by the 12 tools above. The database is untouched — only the tool names/shapes changed, not the storage or graph model.

Search (like a search engine)

search_memory no longer needs a single keyword. It tokenises your query and ranks memories by relevance, so full sentences work:

• Full-text + BM25 — a SQLite FTS5 index over content/tags/category/ details (kept in sync by triggers), ranked with BM25. Multi-word / long queries match memories containing any (or, with mode="all", every) term, with prefix + Porter stemming (sync matches syncing).
• Graph spreading activation — with expand=True (default), memories connected in the knowledge graph to a text hit are pulled in with a decayed score, so related context surfaces even without the query words.
• Ranking blends text relevance with importance and recency. Each result carries relevance (0..1), match_type (text | related | list), matched_terms, and distance (hops from a text hit).
• Fallback — where a SQLite build lacks FTS5, search degrades to a tokenised LIKE term-coverage scorer, so it always works. summarize_memories reports the active engine (fts5-bm25 or like-fallback).

Example: search_memory("Burp Firefox proxy sync") returns the relevant memories ranked, plus anything linked to them — in a single call.

Install

From PyPI:

python3 -m pip install brainmemory-mcp

From a local checkout:

python3 -m pip install .

Both install the package and a console script named brainmemory-mcp.

For development (editable install):

python3 -m pip install -e ".[dev]"

To build/publish a release, see docs/RELEASING.md.

Run

stdio mode (default)

Best for local MCP clients that launch the server themselves. Memory in ~/.brainmemory-mcp.

brainmemory-mcp

# Or without the console script
python3 -m brainmemory_mcp

# With a custom memory location
brainmemory-mcp --data-dir /path/to/memory

Web mode (HTTP + SSE)

Enable with --web for remote / networked clients.

# Defaults: 127.0.0.1:8765, memory in ~/.brainmemory-mcp
brainmemory-mcp --web

# Custom host/port and memory location
brainmemory-mcp --web --host 0.0.0.0 --port 9000 --data-dir /path/to/memory

Endpoints once running in web mode:

• SSE stream: http://<host>:<port>/sse
• Message POST: http://<host>:<port>/messages/

Configuration

Connect a client

stdio (recommended for local use)

Configure the client to launch the server as a subprocess:

{
  "mcpServers": {
    "brainmemory": {
      "command": "uvx",
      "args": ["brainmemory-mcp"]
    }
  }
}

If installed on your PATH, you can use "command": "brainmemory-mcp" with "args": [] instead.

Web (SSE)

Start the server with --web, then point an SSE-capable client at the /sse endpoint:

{
  "mcpServers": {
    "brainmemory": {
      "url": "http://127.0.0.1:8765/sse"
    }
  }
}

How memory is stored

Memories live in ~/.brainmemory-mcp/memory.db (SQLite, WAL mode) across three tables:

• memories — nodes: id, content, category, tags, importance (1–5), created_at, updated_at.
• memory_details — extra facts attached to a memory (cascade-deleted with it).
• memory_links — directed connections source_id -> target_id with a relation and weight (cascade-deleted with either endpoint).

Search uses a SQLite FTS5 full-text index (memories_fts, kept in sync by triggers) ranked with BM25, augmented by graph spreading activation. Graph operations (multi-hop recall_related, shortest-path connect_memories, degree centrality in summarize_memories) are computed with plain SQL + a little Python — no external services or vector database required.

Nothing is ever silently deleted — removal only happens through forget_memories or explicit delete ops in edit_details / edit_links. When an older database is opened that lacks the newest schema (the graph tables or the FTS index), it is backed up automatically to ~/.brainmemory-mcp/backups/ before the new objects are added.

License

MIT