← AI Terminology
Agent Memory
Agent memory is the store of information an agent retains across steps or sessions — short-term working context, episodic logs, and long-term retrieved knowledge.
It goes beyond the raw context window with external memory systems.
It goes beyond the raw context window with external memory systems.
Why It Matters in AI
Context windows are finite and sessions reset. Memory layers let agents remember user preferences, past tools results, and project state. Designing what to store, summarise, and retrieve is a core agent-architecture problem.
Key Points
| Aspect | Description |
|---|---|
| Ops | Summarisation, decay, privacy, user edit/delete |
| Related | RAG, context engineering, MCP resources |
| Episodic | Timestamped interaction logs |
| Semantic | Distilled facts and preferences |
| Long-term | Vector DB, SQL, knowledge graph of past events |
| Short-term | Current scratchpad / conversation buffer |
Simple Analogy
A personal assistant with a notebook: today’s meeting notes on the desk (short-term) and a filing cabinet of past clients (long-term) they can look up.
Common Usage Examples
- MemGPT/Letta-style memory tiers
- LangGraph checkpointers
- User profile stores in chat products
- Retrieve last sprint decisions for a coding agent
Summary
In short: Agent memory persists useful state beyond a single prompt — short-term scratchpads plus long-term stores the agent can recall.