Agent Memory: An Anatomy
4 hours ago
- #Agent Memory
- #AI Design
- #Cognitive Science
- Agent memory libraries use cognitive science terms like episodic, semantic, and procedural memory, but their implementations often fall short of these concepts.
- Most 'agent memory' systems are actually narrower autobiographical memory systems, storing facts about the user's life and preferences on their behalf.
- An agent memory system consists of three core components: an extractor (compresses conversations into facts), a store (database for statements), and a retriever (fetches relevant statements).
- Key design choices include extraction timing (eager vs. lazy), handling contradictions (overwrite, append, or supersede), and retrieval methods (vector search, keyword filters, reranking).
- The four memory types are episodic (events), semantic (facts), procedural (skills), and prospective (future intentions), but production libraries mainly handle semantic memory.
- Biological analogies (e.g., consolidation, emotional salience, forgetting) are useful for vocabulary but can be misleading as design guides, as agent systems have different constraints and capabilities.
- Consolidation (offline reorganization of memories) is a valuable feature to import, while emotional salience is largely absent in text-only agents, and biological-style forgetting may not be desirable.
- The gap between terminology and engineering is clearest in procedural memory, where labels often mask identical implementations to semantic memory.
- Prospective memory (remembering future actions) is mostly absent in production libraries, representing open territory for development.
- Understanding the components and memory types provides a map to evaluate any agent memory library, regardless of its specific implementation or marketing claims.