Agent memory
Memory is selected context from prior turns or external stores that helps an agent continue work.
In plain terms
Memory is the agent's notebook. It can include user preferences, task state, summaries, or retrieved facts.
Why it matters
Long tasks need continuity, but full transcripts are expensive and noisy.
How it works
Store durable facts separately from temporary scratch state. Retrieve only what is relevant to the current task.
When you use it
Use memory for recurring users, long projects, and assistants that need preferences or history.
Common mistakes
- Saving everything forever.
- Mixing untrusted content with instructions.
- Letting stale memory override current user intent.
Best practices
- Make memory inspectable and deletable.
- Summarize old context.
- Tag memory by source and confidence.
Try it yourself
Add a tiny preferences store to a study agent and let the user view or delete saved facts.