How agents remember

Talk to a public chatbot today and it has forgotten you by tomorrow. Talk to an EnClaw agent and something different happens: mention in March that you prefer figures in tables, and in June the figures arrive in a table. Agents remember — deliberately, selectively, and with rules about who owns what.

After a conversation ends, the agent distils it: what was asked, what was done, what's still pending, and anything worth keeping about how you like to work. The transcript itself isn't hoarded forever; the useful residue is. Next time you talk, that residue is quietly in the room — which is why the second week of working with an agent feels noticeably smoother than the first.

Whose memory is it?

Not everything an agent learns belongs to everyone, and EnClaw keeps the layers apart.

Some things are organisation-level: the financial year ends in June, the product is called Atlas, invoices go through Priya. Facts like these are useful to everyone, so they're shared — teach an agent once and the whole organisation stops repeating itself.

Other things are personal: your writing style, your priorities, the fact that you want the blunt version first. Those stay attached to you alone. A colleague chatting with the very same org agent doesn't inherit your preferences, and can't fish your details out of it either. Your conversation history is equally your own — each person sees only their own sessions.

One drawer per project

Memory is also kept apart by place. Each project — and each Slack, Teams or Telegram channel, automatically — is its own separate context. What the agent learns while working in the #acme-renewal channel stays filed under that project; ask it something in #internal-ops and it doesn't drag Acme's numbers into the answer. Think of it as one drawer per client: nothing leaks between drawers by construction. You don't set this up; it simply happens.

You hold the eraser

Remembering is only trustworthy if forgetting works too. If an agent has something wrong — an old job title, a stale figure, a preference you've changed — correct it in plain words: "that's out of date, the renewal moved to October." Corrections don't just fix the moment; they shape what's kept, so the same mistake stops recurring. And if you want something gone outright, say so: "forget what I told you about the Acme budget." The practical details — what to say and what to expect — live in How memory works day to day.

So agents can act, and they remember with manners. That leaves the question underpinning both: when an agent is about to do something that matters — send the email, create the ticket, spend the money — what stops it doing the wrong thing? That safety net is the next chapter, and it's the reason your organisation trusts any of this.