Memory & context
Good colleagues remember things. EnClaw agents do too: facts you've shared, preferences you've expressed, corrections you've made and the thread of recent work — all carried forward so you don't start from scratch every conversation.
This page explains what gets remembered, who can see what, and how to steer it.
What agents remember between conversations
After each conversation, the agent quietly takes stock: what you were trying to do, what got done, what's still pending, and anything worth keeping — a preference ("Max likes bullet points"), a fact ("the financial year starts in July"), a follow-up ("chase the Acme contract next week").
The next time you chat, the agent brings the relevant bits back with it. That's why you can say "same format as last week's report, please" and it knows what you mean, or why it stops offering afternoon meeting slots after you've mentioned you prefer mornings.
Memories aren't a verbatim recording of everything you've ever typed — they're the useful distillate, and they naturally fade: a preference you mentioned once, months ago, carries less weight than one you've repeated recently. Recent, repeated and corrected information wins.
Personal vs organisational memory
Not everything belongs in the same drawer. EnClaw keeps memories in separate, scoped spaces:
| Scope | Example | Who it serves |
|---|---|---|
| Organisational | "Our standard proposal template is the two-pager" | Everyone — shared working knowledge |
| Team | "The delivery team's standup is 9:15" | That team |
| Agent | "This agent always reports values in AUD" | Everyone using that agent |
| Personal | "Max prefers morning meetings and hates jargon" | You only |
When you tell an agent something, it judges where it belongs: a fact about how the company works becomes shared knowledge; a fact about you becomes personal. Your personal memories are private to you — no colleague's conversation can surface them, and another user asking the same agent the same question won't get answers flavoured by your preferences or your data.
The same wall applies to conversation history: each person sees only their own past sessions.
Per-project context — channels keep to themselves
Here's a quiet superpower: every Slack or Teams channel is automatically its own isolated project context. No setup needed.
An agent working with you in #project-acme builds up context about Acme — the milestones, the people, the decisions. Ask the same agent something in #project-zephyr and it works from Zephyr's context instead. Nothing bleeds across: Acme's confidential pricing discussion will never surface in the Zephyr channel, by design.
Practical upshots:
- Put project talk in the project's channel. The channel is the filing cabinet.
- Don't be surprised if an agent in one channel doesn't volunteer something you told it in another — that's the isolation working. If you want it available in both places, mention it in both.
- Direct messages have their own context too, separate from any channel.
Correcting an agent — and why it sticks
When an agent gets something wrong, tell it, right there in the conversation:
"No — the report should cover the calendar month, not the last 30 days."
"Actually, Priya owns that account now, not James."
Two things happen. The agent fixes the answer at hand, and it records the correction so future answers are shaped by it. Corrections are the highest-value thing you can give an agent — a couple of weeks of casual "actually, it's like this" remarks and it starts anticipating what you want. If it held a contradictory memory, the correction wins and the old version is retired.
You don't need any special wording. Correct it the way you'd correct a colleague.
Asking an agent to forget
Memory is yours to prune. Just ask, in plain words:
"Forget what I told you about the Meridian budget."
"Please forget my home address."
The agent removes the memory, and it stops influencing future answers. Good moments to use this:
- You shared something sensitive that didn't need keeping.
- A fact is now wrong and you'd rather delete it than correct it.
- You're experimenting and don't want a throwaway detail to stick.
If you'd like something never remembered in the first place, say so up front — "don't remember this" — before sharing it.
What memory is not
- Not surveillance. Memory serves your conversations; it isn't a feed anyone browses to see what you've been asking.
- Not permanent. Retention rules set by your organisation apply, memories fade with time, and you can always ask for something to be forgotten.
- Not shared by default. Personal stays personal; only genuinely organisational knowledge is shared, and each project channel keeps its own counsel.
If an agent seems to have forgotten something you're sure you told it, the usual culprit is context isolation — you told it in a different channel. See Troubleshooting for the quick fixes.