The upstream Memory server keeps everything in one place; here the file path travels with the call, so work notes and personal notes stay apart. Getting the benefit means the model has to pass that path every time — which is why the README ships a system prompt rather than leaving it to chance.
A knowledge-graph memory server built on the MCP team's Memory server, with two things added: the memory file path can be set per call, so each project keeps its own graph, and the server can produce timestamps to attach to what it stores.
- Entities, relations and observations created, read and deleted across a persistent knowledge graph
- The whole graph read back, searched by term, or opened at specific nodes
- A memory file path chosen per project rather than one shared store — pass it with each operation
- A current timestamp on demand, so entities and observations can record when they were written — `get_current_time`
- A system prompt in the README that tells the model when to retrieve, what to record (identity, behaviours, preferences, goals, relationships) and when to update
Node.js version 16 or higher. No account and no key. Run the built `dist/index.js` as a Node process. `MEMORY_FILE_PATH` sets the default store; without it the server writes `memory.json` next to the script. Installable into Claude Desktop through Smithery. The package is `memory-custom` 1.0.0.
One command — npx -y github:BRO3886/mcp-memory-custom