This page is the product-level map for SuperLocalMemory V3.7. It separates what the runtime provides from what an operator must explicitly configure or verify.

SLM stores local memory as atomic facts, episodic scenes, temporal events, canonical entities, profiles and scopes. A write creates a durable operation: raw → queryable → enriching → complete, with retryable failures retaining their evidence. slm remember --sync waits for declared materialization work; the default write receipt is intentionally queryable-first.

Recall can combine dense semantic, BM25 lexical, temporal, Hopfield associative, and spreading-activation candidate channels. It then applies fusion, optional reranking, and graph-derived score enhancement. Use slm trace "query" to inspect channel participation and evidence. SLM does not represent its internal ranking utility as calibrated answer confidence.

The local Brain contains consolidation, behavioral patterns, feedback and outcome records, rewards, soft prompts, and ranking-related learning components. The Entity Explorer surfaces compiled entity summaries and timelines. Skill Evolution maintains lineage, budgets and verification outcomes for opt-in evolution workflows.

These components learn from qualified evidence and configured feedback. They do not imply that every observed result was good, that a skill will improve, or that an agent may execute a generated change without review.

SQLite plus sqlite-vec are canonical. CozoDB and LanceDB are packaged derived projections for graph and vector work. They must be activated deliberately:

slm db scale status
slm db scale prepare
slm db scale verify <stage-id>
slm db scale promote <stage-id>
slm db scale rollback <backup-id>

The lifecycle uses the expand–verify–promote–rollback pattern. It prevents an installed backend from silently becoming authoritative and gives the operator a named rollback point. SLM does not publish an unqualified memory count, latency, or throughput guarantee for this feature.

Optimize has three surfaces:

Exact cache, tags and invalidation are the stable cache path. Safe compression uses conservative normalization; aggressive prose compression is opt-in and lossy. MCP and skill use do not intercept the primary conversation turn unless a proxy is in the path.

SLM Mesh provides authenticated messages, queues, locks, inbox/outbox, and optional peer discovery. It coordinates configured peers; it is not a replicated, conflict-resolving memory database.

SLM exposes CLI, Python, MCP HTTP/stdio, dashboard, a Claude Code plugin, and an additive Codex add-on. Named IDE configurations are available through slm connect; hooks, IDE configuration, connectors, cloud backup, and Gmail, Calendar, or transcript adapters must be enabled explicitly. Dynamic memories are retrieved at runtime as bounded untrusted evidence, not copied into IDE instruction files.

SLM includes provenance, policy, audit, lifecycle, export/erasure, health, diagnostic and backup surfaces. These controls can support a deployment’s governance work; they are not legal certification.

Before relying on a deployment, run:

slm doctor
slm health
slm status
slm trace "a representative production query"

Then verify the selected MCP client, adapters, provider and network boundary in the environment that will actually run them.