Search the memory system by semantic similarity. Use this at the start of any topic-specific conversation to surface relevant context, prior decisions, and evolved thinking — without being asked. Returns memories ranked by relevance × salience.

Save a new memory. Use this to record decisions made, insights surfaced, how the user's thinking has evolved, feedback given, or important context from this conversation.

Get a synthesized brief on a topic — what is known, how thinking has evolved, and what open questions remain. Use before deep-diving into any recurring topic like web onboarding tests, the stock portfolio, or an M&A situation.

Given a memory ID, find semantically related memories. Implements spreading activation.

Run the consolidation job: extract semantic patterns from recent episodic memories. Run this at the end of sessions covering important topics. This is what makes the memory system get smarter over time.

List recent memories, optionally filtered by type.

Return total memory count and breakdown by type.

Inspect a single memory by ID: full content plus all metadata (type, topic, salience, timestamps, access history, source). Use before editing or deleting to see exactly what's stored.

Edit a memory in place — fix wrong content, adjust salience, retag topic, or reclassify type. Only the fields you pass change; id, created_at, and access history are preserved. Content changes are re-embedded automatically.

Permanently delete a single memory by ID. Irreversible — inspect it with get_memory first, and confirm with the user before deleting anything they didn't explicitly ask to remove.

Export memories to a JSONL file (one memory per line, embeddings excluded). Use for backup before risky operations, or to move memories between machines.

Report near-duplicate memory pairs by embedding similarity. Report-only: nothing is deleted. Review the pairs, then remove specific losers with delete_memory after confirming with the user — prefer deleting the lower-salience or older duplicate.

Log a forward-looking assessment to the judgment ledger. Use this whenever making a prediction, recommendation, or forward-looking call. Prefer a numeric 'probability' (0-1) — it enables real calibration (Brier score) over time. Assessments are tracked and later scored against what actually happened.

List unresolved assessments from the judgment ledger. Use during weekly review to surface what needs scoring, or at session start to remind of open calls in a domain.

Mark an assessment as resolved with its actual outcome and score. score: 1 = right, 0 = partially right, -1 = wrong. Call this as soon as the outcome of a prediction is known.

Extract calibration patterns from resolved assessments in a domain and write a high-salience feedback memory. Run this after a batch of resolutions in a domain. Requires at least 3 resolved assessments in the domain.

Generate a structured bias report across all resolved assessments. Shows where judgment is well-calibrated vs. systematically off. Use for quarterly self-calibration review.