Ninety-four tools would swamp a context window, and most memory servers make you take all of theirs. Here the profile is a single environment variable: a coding agent gets the 31-tool `code` surface with the graph tools, a chat client gets the 16-tool `core` one, and neither pays for the other's. The scoring contract is the other unusual piece — it refuses to report an answer confidence it cannot compute, and says so in the response.
SuperLocalMemory (SLM) keeps agent memory on your own machine in SQLite with sqlite-vec, and exposes it over MCP, a CLI, a dashboard and IDE integrations. Recall is not one vector lookup: semantic, BM25, temporal, Hopfield and spreading-activation candidate channels are fused, and every memory carries ingestion timing and provenance so an agent can ask what changed as well as what is true.
- A profile switch that decides how much tool surface reaches your client: `core` 16 tools, `code` 31, `mesh` 8, `full` 49, `power` 61, `whole` 94 — set through `SLM_MCP_PROFILE`, and an unknown name stops startup rather than silently selecting another set
- Writes that return a durable receipt with `operation_id`, `fact_ids` and `materialization_state`, plus an `idempotency_key` so a retry is not a duplicate — `remember`
- Natural-language recall whose scores are explicitly separated: `relevance_score` for query relevance, `ranking_score` for diagnostic ranking, `memory_confidence` for the stored assertion, and `answer_confidence` left null because retrieval scores are not answer probabilities — `recall`, `search`, `fetch`
- Scoped memory — `personal`, `shared` with named profiles, or `global` — with cross-profile recall default-deny
- Code-graph tools for review and change work: blast radius, review context, architecture overview, dead-weight hunting — `build_code_graph`, `get_blast_radius`, `get_review_context`, `get_architecture_overview`, `find_large_functions`
- Explicit context optimisation rather than silent compression — `slm_compress`, `slm_cache_set`, `slm_cache_get`, `slm_optimize_stats`
- Lifecycle and governance surfaces: retention policy, consolidation, audit trail, health and a recall trace that shows which channels contributed — `set_retention_policy`, `compact_memories`, `audit_trail`, `recall_trace`, `consistency_check`
- Correction handling that keeps a human in the loop — `correct_pattern`, `review_correction`, `list_corrections`
Python 3.11 or higher, into a virtual environment: `python -m pip install superlocalmemory` gives you the `slm` CLI and the importable SDK. Two transports expose the same tools — HTTP against one shared daemon at `http://127.0.0.1:8765/mcp/`, or stdio via `slm mcp`. The default local runtime needs no Docker, no separate graph database and no API key; provider-backed enrichment, cloud backup and connectors are separate choices you make. Licensed AGPL v3.
One command — npm install -g superlocalmemory