What agent memory means
Agent memory is not a longer prompt. It is a separate place your agent stores what it learns, and reads from when it needs to.
Instead of cramming a whole history into every request, the agent works with a store it can write to and read from:
- Write: keep what is worth remembering from an interaction, so it survives the session.
- Read back: pull the pieces relevant to the current task, instead of carrying everything.
How systems do this varies. Some keep raw text; others distill it into shorter notes or facts. Some find memories by keyword; others by meaning. Simpler setups paste what they find straight into the prompt, while more capable ones let the agent reason over it before answering. What stays constant is the shape: a durable store the agent manages on purpose, separate from the prompt, that holds up across sessions and machines because it lives outside any single conversation.
Takeaways
- Memory is a separate store the agent writes to and reads from, not a bigger prompt.
- The core loop is simple: write what matters, read back what is relevant.
- Approaches differ in what they store, how they search, and how much they reason over results.
Next: What you can build with memory.
Give your agent long-term memory in minutes.
Try Hindsight free