An AI agent with persistent memory

Most agents wake up with amnesia: every task starts from zero, in a fresh sandbox, with no memory of you. Here is what persistence actually has to cover for an agent to improve over time — and the controls you need once it does.

"Memory" is one of the most-used and least-defined words in AI products. Sometimes it means the chat history in the current window. Sometimes it means a few facts the product saved about you. Almost never does it mean what people picture: an assistant that learned how you work last month and is still that assistant today.

For an agent that does real jobs, persistence is not one thing but three, and products differ sharply in which ones they actually provide.

The three layers of persistence

1. Memory of you

What you prefer, how you phrase things, which supplier you already rejected, that "the report" means the Tuesday one. This is the layer most products mean, and it is genuinely useful: an agent that remembers your standards does not need them re-explained. On MyClawn this memory persists across sessions and is classified by visibility — private to you, shareable, or public — so the agent knows not just what it remembers but who it may say it to.

2. A machine that persists

This is the layer almost everyone skips, and the one that matters most for work. An agent that runs in a sandbox created per task loses every file it made, every login it established and every tool it installed the moment the task ends. Next time, it starts in an empty room. An agent with a computer of its own keeps all of it: the spreadsheet it built last week is still on the desk, the portal is still logged in, the app it set up is still installed. Memory ofyou tells it what you want; a persistent machine lets it build on what it already did. See what giving an agent its own computer involves.

3. Memory of the work

What it sent, what it changed, what happened last time. This is the accountability layer: the machine's own record of the agent's actions — the sent folder, the files, the history — as opposed to the agent's narrative of them. When the agent says it already emailed the supplier, you can check the sent mail rather than take its word.

The controls persistence makes necessary

The moment an agent remembers, three questions become urgent, and a product that remembers without answering them is a liability.

• Can I turn it off for a moment? Some tasks should leave no trace — a sensitive question, a one-off you do not want influencing future behaviour. MyClawn's private mode makes a turn ephemeral: no memory is read, none is written, and the session is destroyed afterwards.
• Can I remove something specific? Memory that is wrong persists just as well as memory that is right. You need surgical forgetting — delete the entries that match, and get a receipt of what was removed — not a single "clear everything" button.
• Who else can see it? Persistent memory that leaks into places it should not is worse than no memory. The visibility classification exists so the boundary is enforced mechanically, not by hoping the model is discreet. The guarantees, as implemented, are on the Security & trust page.

What persistence changes in practice

The first task an agent does for you is always the most expensive: you explain the context, it finds the material, it learns the shape of the job. With persistence, that cost is paid once. The tenth supplier search reuses the sheet and the standards from the first nine. The weekly portal check needs no login. The agent gets cheaper to instruct and more accurate the longer it runs — which is the only version of an agent that is worth keeping.

What still breaks

• Stale memory. The supplier you rejected in March may be the right one in September. Review what the agent remembers about anything that changes; forget what is out of date.
• Confident wrong recall. An agent can remember a preference you never stated. Watch its first few uses of any remembered fact — the desktop is streamed live, so you can see it act on what it thinks it knows.
• Machines fill up. A persistent computer accumulates files like any other. Give it a tidy-up task occasionally, the same as you would a person's desk.

Where this is the wrong tool

For a one-off question, memory is overhead. For anything you will ask again — with the same context, the same documents, the same logins — it is the whole point. How the memory model, private mode and the clone's machine fit together is described in Cloud clones.