Memory & Learning
What the agent carries forward — and what it deliberately doesn't.
By default, an agent starts every conversation knowing nothing about who it is talking to. That is often correct: an agent processing loan documents doesn't need to remember anyone.
But some agents get better by remembering, and Qore separates that into distinct kinds — because they have different scopes, different risks, and different rules about what may enter them.
Three kinds of memory
User persona memory
What an individual user has asked about, preferred, or corrected
Off
FAQ memory
Canonical answers to questions the agent already knows
On
Learning cases
Reviewed corrections from real production behavior
Manual approval only
User persona memory remembers what an individual user has asked about, preferred, or corrected — across conversations. Turn it on for assistant-style agents where personalization helps. Leave it off for agents handling discrete, stateless tasks, and for headless agents called by API with no persistent user identity: there, it adds noise without benefit.
FAQ memory lets the agent reuse a stored canonical answer when a question matches a known FAQ, rather than answering from scratch each time. On by default, and the right default for support-style agents.
Check the scope before you enable memory. Whether persona memory is confined to one agent or shared more widely determines who can see what a user told which agent — a question worth answering explicitly in a regulated environment. Confirm the current behavior with your platform administrator before turning it on for an agent handling sensitive material.
Learning cases are corrections captured from real production behavior — a case where the agent got something wrong and a person recorded what the right answer would have been.
Why learning cases require a human
Learning cases do not enter memory automatically. Each one is reviewed and approved by a person first.
Auto-approval is intentionally disabled. This is a deliberate design decision, not a missing feature. An agent that learns automatically from its own traffic can be taught the wrong thing by a bad input — accidentally by an unusual case, or deliberately by someone who works out that it learns. Manual approval is the boundary that makes memory a safe place to write to.
This is the clearest example of a principle that runs through the platform: the closer a mechanism gets to changing the agent's future behavior, the more a person has to be in the loop. See Human-in-the-Loop.
The improvement loop
Memory is where the operating loop closes.
Monitor production behavior and score live conversations to find where the agent missed.
Record what the correct behavior should have been — a learning case.
Review and approve it. This is the gate.
Validate the improvement with Evals before rolling it out more broadly.
The agent improves from real outcomes, not just from isolated test runs — but only through a door a person opens.
Context is not memory
A related idea, easy to confuse. Memory is what the agent carries across conversations. Context is what it holds within one — the conversation so far, retrieved documents, tool results.
Context is finite. Long conversations eventually exceed what the model can hold, which is what context compaction addresses. Retrieval — pulling only the relevant material into context rather than everything — is covered in Knowledge & Context.
Where this is configured
Agent Behavior — the Memory and FAQ memory toggles
Reinforcement Learnings — the learning-case approval flow
Monitoring and Online-Eval — where misses are found
Evals — validating an improvement before it ships
Last updated