Store memory in 1 of 6 layers for an entity. Rejects pasted assistant output / CI logs unless `force=true`. Set `importance=1.0` to pin (survives auto-forget).
recall
FTS5 + heat × momentum × importance composite ranking with `match_reasons` explaining WHY each row matched. Supports pagination (`offset`/`has_more`), `band` filter, layer aliases (`decisions`/`warnings`/`how`/...), and `mark_accessed=false` for passive previews.
recall_file
Complete edit history of a file across all sessions, with per-edit user-intent context. **v0.3.0** `scope_to_roots` flag filters by client roots.
update_memory
**v0.1.0** Atomic edit of an existing memory. Preserves `memory_id` (session_file_edits links stay intact). Prefer over forget+remember.
list_entities
**v0.1.0** List what the memory knows about — cheapest "what do I know?" primitive. Filter by `kind`/`min_memories`; returns layer breakdown per entity.
read_smart
Diff-only file read. Returns full content on first read, ~50 tokens on unchanged re-reads, only changed chunks on real edits.
forget
Explicit delete OR auto-sweep based on `forgettingRisk`. Pinned (`importance>=1.0`) and caveat-layer memories are always preserved. **v0.3.0** `interactive` flag asks the user via Elicitation before deleting a specific memory_id.
consolidate
Sleep-mode compression: cluster cold low-importance memories → protected learning-layer summary. Supports `dry_run` preview. **v0.3.0** `use_llm` flag asks the client LLM (Sampling) to rewrite cluster summaries into prose.