Evidence-first authority page
AI Agent Memory Types: Short-Term, Long-Term & Episodic
The three distinct memory types an agent architecture can implement, and why conflating them leads to unreliable agent behavior.
Direct answer
AI agents can implement three distinct memory types — short-term (the current conversation's context window), long-term (persisted facts retrieved via a vector store or database across sessions), and episodic (a record of past specific interactions or actions) — and each requires a different storage and retrieval mechanism, so an agent architecture should be explicit about which type it's actually implementing rather than treating 'memory' as one undifferentiated feature.
Last reviewed 2026-08-29
This page remains noindex while its page-level source receipts are completed and verified. It is available for review, not presented as publication-ready authority.
Short-term memory: the context window itself
The simplest form of agent memory is just the current context window — everything said in the current session, held in the model's active context. This requires no additional infrastructure but is bounded by the context window size and disappears when the session ends.
Long-term memory: persisted, retrievable facts
Long-term memory typically means storing facts (user preferences, prior decisions, domain knowledge) in a vector database or structured store, retrieved via similarity search or lookup when relevant to the current task — this is architecturally the same pattern as RAG, applied to the agent's own accumulated knowledge rather than an external document corpus.
Episodic memory: a log of what happened
Episodic memory records specific past interactions or actions (what the agent did, when, and why) rather than distilled facts — useful for an agent that needs to reference 'what did I already try' or avoid repeating a failed approach, distinct from long-term factual memory which stores conclusions rather than the history of how they were reached.
Why the distinction matters practically
An agent architecture that conflates these — storing everything in one undifferentiated 'memory' store — tends to retrieve irrelevant context (mixing a stale fact with a recent action log) and degrade in quality as the store grows. Being explicit about which memory type a given piece of information belongs to keeps retrieval targeted and relevant.