Not by themselves: a model forgets everything between calls. What looks like memory is data stored outside the model, such as case notes, task records and customer preferences, which the platform loads into the agent's context window at the start of each task. What gets stored, for how long and for whom is a decision people make.
Where does an agent's memory actually live?
In three places, each with a different purpose.
The task record. Steps taken, outputs, what the task is waiting for. This is what lets a task pause and resume, covered in how AI agents handle long-running tasks.
Case and customer history. Previous contacts, outcomes and stated preferences, usually held in your existing systems such as a CRM or case management tool. The agent reads them like any other source.
Agent notes. Things the agent has learned about how to do its job for this business: that a particular supplier always sends invoices in two parts, or that a team prefers summaries in a set order. These are written by the agent, so they need the most care.
At the start of a task, the platform selects from these and places the relevant parts into the context window, the working space the model reads. Choosing what goes in, and in what form, is the discipline known as context engineering.
What should an agent be allowed to remember?
Memory is data, and stored personal data carries the same obligations whether a person or an agent wrote it. A few rules hold up well:
- Remember facts with their source. "Customer prefers email" should record that the customer said so, in which contact, on what date.
- Do not let memory upgrade a fact. Something a customer claimed stays a claim when recalled later. It does not become verified by being remembered.
- Scope memory by customer and purpose. Notes made in one customer's case must not surface in another's.
- Set expiry. A preference from three years ago may no longer hold. A fraud flag must follow its own retention rules.
- Make it erasable. If a person exercises a right to erasure, memory stores are in scope like any other.
What may the agent do, and what stays with a person?
| Memory activity | Agent | Person |
|---|---|---|
| Read case history permitted by its role | Yes | Sets the role's scope |
| Write task progress | Yes | None |
| Write notes about how to do the job | Yes, marked as agent-written | Reviews notes that change behaviour |
| Record sensitive facts, such as vulnerability or health | Only where the role permits and the source is recorded | Defines what may be kept |
| Delete or correct memory | On instruction | Decides on disputed corrections |
What does this look like in practice, and what goes wrong?
Take a finance team's agent that helps with month-end close. Each month it learns small things: one subsidiary always posts its intercompany charges late; a particular accrual is reversed on day two, not day one. It writes these as notes, each tied to the month and the evidence.
Next month, when the close begins, the platform loads the relevant notes into the agent's context. The agent chases the late subsidiary on day one rather than waiting. A controller can see the notes, correct one that has become wrong, and delete one that was never right. The tasks themselves follow the month-end close checklist, and the notes make each run a little better without changing the model.
Three failures recur. Notes accumulate until the context is crowded with stale guidance. A wrong note persists and repeats an error every month. And memory written during one agent's task feeds another agent's reasoning, so a misreading in one place spreads.
That last one also drives loops: two agents can keep re-reading each other's notes and re-opening the same issue, a pattern addressed in how to prevent loops between AI agents.
What does the platform layer have to do?
- Store memory outside the model, with source, author and date on each entry.
- Load only what the current task and role permit.
- Keep agent-written notes distinguishable from facts read from systems.
- Expire and erase on schedule and on request.
Where Pryme fits
In Pryme, qualification changes an agent's brief, sources, bounds and certification, never model weights. What an agent carries between tasks is held as data with its provenance stamped, so a remembered external claim stays external. The wider design is in the governed agentic operating system guide.
Questions people ask
Is a larger context window the same as better memory?
No. A larger window lets more fit into a single task. Memory between tasks still has to be stored, selected and loaded, and loading too much makes the agent worse, not better.
Does the model learn from our data over time?
Not unless someone retrains it. In a governed setup, improvement comes from changing the brief, sources and notes, which you can inspect and reverse.
Can a customer ask what an agent remembers about them?
Yes. Memory holding personal data falls within data protection rights, including access and erasure, so it must be findable by customer.