Memory by use case
AI memory for research agents
Long investigations that survive a full context window.
Why memory matters here
A research agent may read hundreds of pages over hours. Its context fills up long before the work is done, and anything that falls out is lost unless it was written down. Memory here is a notebook: findings, sources, dead ends and the current plan.
Who a memory belongs to
One memory per investigation, shared by every agent or sub-agent working on it.
The moment it shows
After the context is cleared, it picks up the plan at step six and does not search for what it already found.
What to remember
- Findings, each with the source it came from.
- Dead ends, so they are not explored twice.
- The plan, what is done and what is next.
What to forget
- Raw page text once the finding has been extracted.
- Claims that a later source contradicted. Replace, do not pile up.
- Intermediate scratch work.
See recall before you build it
Add a few facts from your own use case and ask a question.
Research agents: common questions
How is research memory different from user memory?
It is about the task, not the person. It lives as long as the investigation and can be archived or deleted when the report is written.
Should sub-agents share one memory?
Usually yes for findings and the plan, so that work is not repeated. Give each write a source and an author so a wrong note can be traced.
Other use cases
Game characters
Characters that remember what the player did to them.
What to rememberCustomer support agents
Stop asking customers to explain the same problem a third time.
What to rememberCoding assistants
Project conventions learned once, applied in every session.
What to rememberPersonal assistants
An assistant that knows your week, your people and your preferences.
What to rememberPack the doko. Ask it anything.
Nine memories, one question, no account. See which facts an agent would carry into its next answer.