Devin Memory

What devin-memory MCP server does

The devin-memory MCP server provides a durable memory store for agent workflows. It is designed to reduce memory poisoning by screening content before storage, recording where entries came from, and separating active memories from proposed or quarantined entries.

Stored facts can include a source session identifier and source row identifier. These fields let a memory be checked against a real Devin sessions.db when one is available. The server also preserves relationships between conflicting facts instead of silently replacing the older entry. Entries can be superseded, retracted, quarantined, or approved as their status changes.

The project is an unofficial community implementation and is not affiliated with Cognition AI.

How it works

The MCP process communicates over stdio and uses the same retain and recall pipeline as the command-line interface. Writes pass through a screening gate. Content that matches secret or injection-oriented patterns can be placed in quarantine rather than becoming active immediately. The screen operation checks the gate without writing an entry.

Recall returns active memories and supports keyword ranking, tags, workspace scope, and result limits. prime creates a bounded context block for a prompt hook, filtering by workspace and machine profile. Proposed memories produced by session extraction remain inactive until they are reviewed, unless the command is run with its documented auto-approval option.

Conflict detection links newer and older entries. A human or operator can then choose to supersede, retract, or quarantine an entry. The quarantine digest provides a read-only audit hash, while compaction reports or applies duplicate cleanup.

Setup and configuration

Python 3.10 or newer is required for the package. The README recommends installing the MCP-enabled package with uv:

uv tool install 'devin-memory[mcp]'

pipx is listed as an alternative. To run the standalone MCP package, install devin-memory-mcp and configure an MCP host with the devin-memory-mcp command and a database path:

{
  "mcpServers": {
    "devin-memory": {
      "command": "devin-memory-mcp",
      "args": ["--db", "/path/to/memory.db"]
    }
  }
}

The README includes configuration examples for Claude Desktop, Cursor, Zed, JetBrains, and Antigravity. Promotion operations are controlled by DEVIN_MEMORY_MCP_PROMOTE=1; without that setting, the server does not register the release and approval promotion tools.

Tools and capabilities

The MCP tool set includes:

• retain, recall, and screen for writing, searching, and testing memory content
• list, prime, and extract for reviewing or preparing context
• approve, release, quarantine, and retract for status changes
• supersede and conflicts for version and contradiction handling
• verify for checking provenance against sessions.db
• compact and quarantine_digest for maintenance and auditing

The CLI also supports exporting active memories to the JSONL shape used by the referenced memory MCP, allowing the active set to be written to a file for compatibility with that format.

Limitations and notes

Session provenance depends on access to Devin's sessions.db; without that database, claimed source sessions cannot be verified. Session extraction is heuristic and produces proposed entries that require review unless auto-approval is selected. Recall returns active entries only, so quarantined, proposed, and retracted memories are not part of normal recall results.

The MCP server uses stdio, so its host must launch the local command and provide the database path. The README does not describe a remote HTTP transport or a built-in multi-user service.