mcp-memory-service
Purpose-built memory for coding agents. It silently watches what your Claude Code or Cursor agent does across sessions — which files it edits, which decisions it makes, which patterns it follows — then consolidates that into retrievable memories. The benchmark claims 86% recall at 5 for session retrieval, which is unusually strong.
- Engineering teams using Claude Code or Cursor daily
- Anyone tired of re-explaining the same architecture to their agent every session
- Teams that want retrieval quality benchmarks, not vibes
What you'll do
mcp-memory-service runs as an MCP server that consolidates your agent's coding sessions into a retrievable memory. Install via uvx (no cloning needed). Budget 10 minutes.
Before you start
- Python 3.10+ with uv installed (`pip install uv`)
- Claude Code, Cursor, or any MCP-speaking agent client
Step-by-step install
- 011. Verify uv is available
uvx is how we run Python MCP servers without installing them globally.
uvx --version
- 022. Add the MCP config
Add mcp-memory-service to your client's MCP config.
{ "mcpServers": { "memory-service": { "command": "uvx", "args": ["mcp-memory-service"] } } } - 033. Restart your client
Fully quit and relaunch. The first time you use a memory tool, uvx downloads mcp-memory-service (takes ~20 seconds).
- 044. Work normally
Code as usual. The service autocapture runs in the background. It extracts project context, architecture decisions, and recurring code patterns without you needing to label anything.
Your first 10 minutes
- 01Code on a real project for 30 minutes as usual.
- 02Start a new session tomorrow. Ask 'what did I work on yesterday and what decisions did I make?' — retrieval should show the previous session's context.
- 03Inspect the memory graph. mcp-memory-service ships with tools to list and edit what it has stored.
- 04Prune anything off-base. The consolidation step occasionally captures noise; pruning early improves quality.
- 05Add Cognition CLO on top to model which project concepts your team is about to forget.
Troubleshooting
uvx command not found.
Install uv: `pip install uv`. Restart your terminal.
My agent isn't retrieving session memories.
Give it an explicit nudge: 'Using mcp-memory-service, recall the context from last session.' After the first few nudges it often starts recalling automatically.
mcp-memory-service holds the knowledge. Cognition CLO models retention per employee per concept using a Weibull forgetting curve — so you see decay before it becomes a missed SOP or a failed audit.