What problem does it solve? Long-running agents need persistent memory, but an undisciplined write path turns memory into noise: overconfident facts, inferred preferences from a single signal, and silent contradictions. This Skill enforces a promotion gate so every write into durable memory is classified, scoped, deduplicated, confidence-scored, and TTL-assigned before it lands. ## Core Features & Use Cases - Four Typed Stores: Separates policy, preference, semantic fact, and episodic summary into distinct tables with distinct retrieval and write rules, plus an append-only events trace for audit. - Promotion Gate: A reference pipeline (classify, scope-check, extract atomic claim, resolve contradictions, assign confidence and TTL) that all writes to memory_fact and memory_preference must pass through. - Contradiction Resolution & TTL: Supersedes old facts instead of deleting them, lowers confidence on ambiguous conflicts, and forces real expiry dates on hedged or speculative observations. - Use Case: A user says "I might use MongoDB for this prototype." Instead of storing "User uses MongoDB" permanently, the gate stores "User is evaluating MongoDB" with confidence 0.55 and a 30-day TTL. ## Quick Start Use the typed-memory skill to evaluate this conversation observation and decide whether it should be promoted into durable memory, with what confidence and TTL.
