Skip to main content
Knowledge is what you write for an agent. Memory is what it learns by working. When memory is on, an assistant carries facts from one conversation into the next: that this customer is on the enterprise plan, that the deployment was moved to Friday, that the contact prefers to be called Alex. Nobody has to repeat it, and nobody has to write it down. The memory panel

​
Two layers

Core memory

A short, standing block that goes into every prompt: persona, standing instructions, decisions taken, profile of who you are talking to. Small on purpose — it is always there, so it must be worth its space.

Long-term memory

Everything else, distilled from conversations into facts and the entities they connect, and retrieved only when relevant. Grows without bound; costs nothing when it is not needed.
The two are independent. An assistant can have a core block and no long-term memory, or the other way round.

​
How it works

1

The turn is stored

After each exchange, the user and assistant text is pushed to the memory service. Never the system prompt, never tool output.
2

It is distilled

In the background, the exchange is turned into facts — short statements with the entities they involve, and when they were true. A later fact about the same relation invalidates the earlier one, so “the budget is 50k” becomes history the moment it becomes 75k.
3

It comes back

At the start of a new conversation, the relevant part of memory is retrieved and put in front of the model; during the conversation, the assistant can search memory itself when a question calls for it.
Recall degrades quietly. If the memory service is slow or unreachable, the conversation goes ahead without it rather than failing — an assistant with amnesia beats an assistant that will not answer.

​
Choosing how it recalls

There is also a choice of what gets distilled: a knowledge graph of facts plus the entities and relations between them — which is what makes multi-hop questions work — or standalone facts, semantically searchable but without entity nodes. The graph is the default.

​
Looking at what it knows

A memory bucket opens onto four views:
Recall test is the fastest way to answer “why did it not remember that?”. If the fact is in the bucket but the query does not surface it, the problem is retrieval, not storage.

​
The part that matters most

Memory is only useful if it remembers the right things for the right person. An assistant serving a hundred customers must never answer one of them with another’s facts. That is decided by the bucket the assistant writes to, and it is worth understanding before switching memory on in production.

Buckets and scope

Who shares a memory with whom — by tenant, by assistant, by project, by person.

Core memory

The always-present block, and how to edit it.