memory-mcp

A persistent, conflict-aware memory MCP server for AI coding assistants (Cursor, Claude Code).

Every AI coding session starts from zero. This MCP server gives the AI a queryable memory store that survives across sessions - and unlike append-only tools, it detects when a new fact contradicts an older one and automatically invalidates the stale memory instead of letting both exist.

How it works

When you store a memory like "we switched from Postgres to SQLite":

1. The content is embedded (OpenAI text-embedding-3-small)

2. Vector search finds similar active memories

3. Claude Haiku checks each candidate - does this contradict the new fact?

4. Confirmed conflicts get invalidated (valid_to = now) - never deleted

5. The new memory is inserted as currently true

6. A supersession link is recorded so the history is traceable

Old facts are never deleted. You can query what was true at any point in time.

Related MCP server: repo-memory-mcp

Tools

Stack

• TypeScript + Node.js 18

• PostgreSQL 16 + pgvector (HNSW index, cosine similarity)

• OpenAI text-embedding-3-small (1536 dims)

• Anthropic Claude Haiku (conflict detection)

• MCP SDK (stdio transport)

• Zod (input validation)

Setup

# 1. Start Postgres
docker compose up -d

# 2. Install dependencies
npm install

# 3. Copy and fill in API keys
cp .env.example .env
# edit .env - add OPENAI_API_KEY and ANTHROPIC_API_KEY

# 4. Run migrations
npm run migrate

# 5. Start the server
npm run dev

Register in Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "developer-memory": {
      "command": "node",
      "args": ["/absolute/path/to/my-memory-mcp/dist/index.js"]
    }
  }
}

Build first: npm run build

Key design decisions

Invalidate, never delete - setting valid_to instead of deleting preserves history. Point-in-time queries (query_history) would be impossible otherwise.

Conflict similarity threshold: 0.45 - the spec suggested 0.75, but measured against labeled contradiction pairs on text-embedding-3-small, true conflicts score 0.50 to 0.82 cosine similarity. At 0.75 the candidate filter silently drops most real conflicts before the LLM sees them. 0.45 catches them all.

RRF over score averaging - cosine similarity and ts_rank live on incomparable scales. Reciprocal Rank Fusion uses only rank positions, so the two search legs fuse cleanly without normalization guesswork.

LLM default to no-conflict - if Claude returns an unparseable response, the code treats it as no conflict. A false negative leaves a stale memory (recoverable). A false positive deletes a true fact (worse).

Project structure

src/
├── index.ts       entry point
├── config.ts      env → typed config
├── types.ts       shared interfaces
├── db.ts          MemoryStore + migrations + vector/keyword search
├── embeddings.ts  OpenAI embeddings client
├── memory.ts      conflict detection + invalidation + ingestion pipeline
├── retrieval.ts   RRF hybrid retrieval
└── server.ts      MCP server + 4 tool handlers