The original memory server keeps everything in a single JSON file that only grows, which becomes a context problem long before it becomes a disk problem. Moving to a document store with versioning is the whole reason this fork exists. Note the strictness on start-up: both path variables are mandatory, and a missing one is a refusal to start rather than a silent default.
A fork of the standard memory server that swaps flat-file storage for a PouchDB document store. Entities, relations and observations work the same way; what changes is that the backing store versions properly and holds up as the graph grows. Every memory is timestamped.
- Graph building: create entities and relations, add observations to an entity that already exists
- Graph reading: `read_graph` returns everything, plus search across nodes and opening named ones
- Removal: delete entities, observations and relations, so a graph that has gone stale can be corrected
- `get_current_time` supplies the timestamp the recommended workflow attaches to observations
- A per-project memory path, so separate projects keep separate graphs instead of sharing one file
- A ready-made system prompt in the README that wires the whole retrieve-attend-update loop for you
Node.js version 16 or higher; PouchDB installs as a dependency. Clone, install, build, and point the client at the compiled entry point; the package is memory-pouchdb, version 1.0.0. Both MEMORY_FILE_PATH and POUCHDB_PATH must be set as absolute paths — the server refuses to start without them. Set DISABLE_MEMORY_FILE to true to use PouchDB alone, and POUCHDB_OPTIONS for extra database configuration.
One command — npx -y @smithery/cli install @bneil/mcp-memory-pouchdb --client claude