Store ONE literal fact (decision, bugfix, convention, gotcha, preference, event). Use explicit entity names and ABSOLUTE dates (never 'yesterday'): the memory model reads literally. Content that looks like instructions to an agent is rejected. scope: 'global' (applies to every project: personal preferences, conventions) or e.g. 'project:'; omitted = this project. pinned=True only when the user says this must always be remembered: pinned notes open every session.

Adaptive recall. Check sufficient/missing: if sufficient is false, the memory may lack the answer, so verify in code/docs rather than assuming. Searches scope (default: this project) plus global; scope='all' searches every project. mode: "auto" (default: lite first, escalates to full when lite finds nothing or the question looks multi-hop/time-related; stop_reason says which ran), "full" (always multi-hop graph expansion: use it when you know the answer joins several facts) or "lite" (vector top-20 + one Jev call that filters them and judges sufficiency: ~4x cheaper than full, no multi-hop).

Pin (or unpin) a note the user wants in every session; SessionStart injects pinned notes first.

List stored memories (newest first) with their top memory types. scope: omitted = every scope.

Delete a memory and its edges.

Merge/promote proposals from consolidation. YOU are the writer: for each item follow instruction, then call memory_resolve(synthesis_id, text), or memory_dismiss if the proposal is wrong.

Store your synthesized fact for a pending proposal. Source notes are kept (merged ones rank lower).

Reject a merge/promote proposal.

Run a consolidation pass now (also runs automatically every 20 writes). all_notes=True re-judges every stored note from scratch (slow: one Jev call per note); use it once after upgrading jevmem.

Node counts, queued (degraded) writes and Jev usage for this server process.

Retry writes queued while Jev was unreachable.