Consolidation, insights & procedures

Three mechanisms that turn accumulation into understanding — all of them proposals.

consolidate — similarity groups on demand

consolidate(project="myapp", threshold=0.86)

Finds groups of near-duplicate active memories by embedding similarity (read-only). Complementary to Sleep — consolidation, proposed: sleep clusters by graph structure, consolidate by pure similarity. Both feed the same action:

merge_memories — the human synthesis

merge_memories(
  source_ids = [45, 46, 51],
  title      = "CEM project — final synthesis",
  content    = "<the synthesis YOU wrote>")

• The synthesis is stored as a new memory with automatic provenance links (Consolida: [[#45]] [[#46]] [[#51]]) — graph edges included.
• Sources become stale + superseded_by the new memory: out of the way in recall, never gone, always recoverable.

Wadachi will never write the synthesis for you silently — the understanding step is deliberately human (or agent-with-approval).

reflect — cross-memory insights

reflect(project=None)

Thinks across memories: surfaces analogies and non-obvious connections that no single memory holds (reusing the entity graph — no LLM cost). Candidates are stored as proposed insights:

• list_insights() — review the queue (also surfaced by get_context — start here, every session);
• accept_insight(id) — promotes it to a real memory linked to its evidence;
• reject_insight(id) — kept on record as rejected, never re-proposed.

review_procedures — from incidents to rules

Recurring incidents (tagged bugfix/drift/anti-pattern…) are clustered by theme; each recurring theme becomes a candidate always-on rule — the brain learning not just facts but how to act. Read-only: it drafts the rule, you decide whether it enters your CLAUDE.md or SCHEMA.md.