The distinguishing choice is that recall works with zero network calls — the default embedder ships with the package, so there is no key to get and no per-token bill for remembering things. The repository marks itself alpha and warns of breaking changes before 1.0, so pin the version you install.
A memory server your MCP client launches locally. Every conversation writes to an append-only store on disk, and the next session can search it semantically. Embeddings run offline from a model bundled with Chroma, so nothing leaves the machine unless you deliberately switch to a cloud embedder.
- Semantic search over everything stored — `recall`, plus `recall_filtered` for a narrowed query and `answer` for text returned with the chunks it came from
- Four ways to write, each with its own shape: `remember` for a free-form observation, `reflect` for a hypothesis-to-result artifact, `anti_pattern` for a looks-right-but-isn't warning, `checkpoint` for working state
- Session continuity — `pulse` reads back the latest checkpoint with reasoning and anti-patterns, `session_close` writes the end-of-session reflection
- Local RAG over your repo — `index_file` for one file, `reindex` to re-scan, `snapshot_index` to persist the live store, `memory_stats` for counts and sources
- `forget` archives instead of deleting, so removed memory is recoverable; `maintenance` and `backfill_epoch` handle the housekeeping
- The MCP guide lists 16 tools, served identically over stdio and over SSE
No account and no key. `pip install "ai-recallworks[mcp]"` (Python 3.11+), then point your client at the `recall-mcp` command — that is the whole config. Set `RECALL_STORE_DIR` if you want the store somewhere other than the default. For a shared network instance instead, run `ghcr.io/recallworks/recall:latest` on port 8787 with an `API_KEY` and talk to it over `/sse` or `POST /tool/{name}`. The package is `ai-recallworks` at 0.5.2; `src/recall/` is MIT, the `enterprise/` tree is BSL 1.1.
One command — pip install "ai-recallworks[mcp]"