Agent memory is the set of mechanisms that let AI agents retain and use information beyond a single model call: working memory (the current context window), epi
Three layers: working memory (the current context window), episodic memory (records of past interactions), and long-term memory (persistent stores, often vector databases, holding facts and preferences). Combining them lets an agent retain useful information across calls.
Because both extremes fail: too little memory and agents repeat questions and lose context; too much and they surface stale facts, drown in irrelevant history, and run up token costs. Production systems curate aggressively, summarizing, scoring relevance, and expiring memory deliberately.