Cognitive science labels—episodic, semantic, procedural—travel straight into agent memory libraries, but the engineering behind them rarely follows suit. 1972 research by Endel Tulving split human memory into event‑based episodic and fact‑based semantic stores, each failing differently. Libraries compress user chatter into flat facts, glossing over the richer episodic context.

The extractor sits at the front line, reading transcripts and deciding what survives. An LLM call, sometimes with a strict prompt, pulls short statements. Timing matters: eager extraction burns tokens on idle chat, while lazy extraction loses coreference cues and temporal anchors. The choice defines how aggressively a library compresses situational data into decontextualized facts.

The store becomes the heart of the system, choosing between vector indices, relational tables, or knowledge graphs. Every statement carries metadata like timestamps and confidence scores, yet the toughest question remains: how to resolve contradictions when a user moves from Paris to Amsterdam. Overwrite, append, or mark superseded—each strategy shapes the user’s remembered reality.

Procedural and prospective memories stay largely missing from current libraries. Procedural knowledge, such as tying shoes, is often mislabeled as semantic, while prospective intent—“send the contract tomorrow”—lacks a true implementation beyond simple scheduled triggers. In practice, most agent memory systems boil down to autobiographical facts, keeping only what the user explicitly shares and what the system deems essential.