The comparison the project draws is with SQLite, and it is the right one: no server to run, no account to create, one file you can back up or delete. That also sets the ceiling — this is a store with good ranking, not an agent framework, and the embedding quality you get depends on which extra you install. Base install falls back to keyword matching with no dependencies at all.
An embeddable memory layer with a built-in MCP server. Memories live in SQLite on your machine and survive session restarts, tool switches and context-window resets, so what you tell one assistant stays available to the next one. The Python API is three calls — `remember()`, `recall()`, `forget()` — and the MCP server exposes the same store to Claude Code, Cursor, Gemini CLI and other MCP clients.
- Cross-session persistence: a decision made Monday is still known on Friday, without replaying the conversation history
- Cross-tool memory — what you teach one assistant is readable from the others pointed at the same store
- Recall ranked by semantic similarity, recency and importance rather than brute-force history search, with automatic categorisation and scope routing (preferences go global, decisions stay in the project)
- Pluggable embeddings: built-in keyword matching with zero dependencies, local sentence-transformers, a local Ollama instance, or OpenAI
- Optional encryption of memory text and metadata at rest, with no external dependencies
- Automatic deduplication that runs during normal use, the way SQLite's auto-vacuum does — nothing to schedule
- Everything local: no cloud calls, no telemetry, no data collection
No account and no key on the base install. `pip install memorymesh`, then point your client at the `memorymesh-mcp` command — the PyPI package is `memorymesh`, published as `io.github.sparkvibe-io/memorymesh`, and the transport is stdio. Extras pull in the optional back ends: `memorymesh[local]` for on-device embeddings, `memorymesh[ollama]`, `memorymesh[openai]`, or `memorymesh[all]`. A hosted instance is also reachable through Smithery with `npx -y @smithery/cli install @sparkvibe-io/memorymesh --client claude` if you would rather not install anything. Runs on Linux, macOS and Windows. MIT licensed.
One command — pip install memorymesh