Product features: agent memory

In user language: what the assistant remembers across conversations, how to write, edit, and clean it. Not the same as a knowledge base (team/personal how-to materials).

Who can use memory

• My agents in a workspace: have personal memory belonging only to your instance of that assistant; switching conversations or chatting days later can still use it.
• Help agent (no workspace selected, or built-in help in Messages): no personal memory; it mainly reads the system knowledge base (this help/ / help/en/ tree) for product how-tos.
• A colleague’s agent: memories are not shared; copying an agent plan does not bring your memory.

You must select a workspace, then in Messages select your own agent — only then does memory apply.

Memory vs knowledge base

Team-shared material goes in the workspace knowledge base; notes only that assistant uses go in the agent knowledge base; “about you, things to remember” go in memory. Employee files, leave, expenses to look up later go in memos and are not injected every round.

Memory layers

Not written to memory: whole chat transcripts, code paths, general notes already in the knowledge base.

Four long-term types (visible on Memory)

How to have the assistant remember

1. Tap Remember on a message (recommended)

In Messages, talking to your workspace agent:

• Your message or the assistant reply title bar has Remember.
• Open it, pick Long-term memory or Daily note, fill title and body (prefilled from the selection).
• After success that message shows Remembered, to avoid writing twice.

Mobile can also use Remember on a bubble; full entry management is Web Agents → My agents → Memory.

2. Say “please remember…” in chat

E.g. “Please remember: later only use APIs written in skills; do not guess paths.”

The assistant should extract the point first, then write memory, and briefly confirm after success. If you actually want to change operating rules (must follow every round), say “write into operating rules”; the assistant updates operating rules instead of only memory.

2b. Train persona / operating rules / user profile in chat

See Practice guides overview. Create-app and vague phrasing: Workspace apps.

3. Auto-extract after chat (on by default)

After each one question, one answer, the system uses AI to decide whether anything is worth keeping long term and writes it — not only when you close the page or switch modules.

• On the agent Overview you can turn off Auto-extract structured memory after a conversation ends.
• In Web Messages conversation menu you can turn on Auto-extract notices (off by default): if this round wrote memory, it prompts “the assistant noted:…”.

4. Auto-archive when chat is too long

When context nears the cap, earlier chat is written to daily notes then compressed so important context is not lost.

How much memory the assistant “sees” in chat

It does not stuff every long-term memory body into the model every time (that would fill context).

Each round roughly includes:

1. Long-term memory list (one line per entry: title + type + one-sentence note)
2. Full text of a few entries most related to the current question (default about 5)
3. Excerpts of today and yesterday’s daily notes (if any)

For other entries the assistant still knows “there was something” from the list; for detail it should search memory, not guess.

To pull original text of earlier chat (what you asked in the last few days, a certain word), tell the current assistant — see Look up past conversations. That is not memory: if you never tapped Remember, memory will not have it, but the conversation record still does.

How the assistant corrects and deletes on its own

Talk directly, e.g.:

• “Last time was wrong, I actually prefer short paragraphs” → the assistant should update that long-term memory
• “Forget the project deadline entry” / “delete that memory” → the assistant should delete it
• “Stop remembering that I like bullets” → same

The assistant searches then edits/deletes per your intent, and briefly confirms. You can also edit or delete entries anytime on the Memory page.

Chat rating “information is wrong”

If you tap Not satisfied on an assistant reply in Messages and tick Information is wrong, the system creates a pending correction on that agent’s Memory page (not written immediately, not auto-injected next round). You can:

• Write to memory: after confirm it becomes a “preferences and corrections” long-term memory
• Ignore: drop this suggestion

Not-satisfied ratings on the help agent go to memory forge (improving help knowledge), not the personal-memory pending list.

Auto-tidy (drop unused / stale)

The system auto-tidies personal memory at night (default a window from about 2am in the server’s configured local time zone):

• Drop exact duplicates and same-named old entries
• Clean expired daily notes (default keep about the last two weeks)
• When there are many entries, a model can review near-duplicates / stale project info then delete

This is not admin-side memory forge (improving help/workspace knowledge): auto-tidy only touches this assistant’s personal memory, not the team knowledge original.

Where you see “what changed”

Open Agent marketplace → My agents → that instance → Memory:

• Auto-tidy log: recent tidies for this assistant (nightly / you ran it / admin site-wide run as it relates to this assistant)
• Tidy my memory: run one round on this assistant now (confirms; may delete duplicate or expired entries)

The admin console has site-wide run history (ops). See Agent memory maintenance (admin).

Where to view and edit

1. Open Agent marketplace → My agents → pick an instance.
2. Open the Memory page:

- Confirm or ignore pending corrections (from chat “Not satisfied · Information is wrong”) - View / edit / delete memory entries - Maintain user profile - View auto-tidy log, or tap Tidy my memory - Toggle auto-extract (and legacy “append whole paragraphs”, not recommended)

1. You can also search memory to try whether a keyword hits.

Web is the most complete; mobile can Remember and see notices; deep edit and tidy logs: open the same agent manage page in a browser.

Relation to conversations

• One conversation = one continuous chat line with an assistant; closing the page is not ending it — you can continue the same conversation.
• New conversation, delete conversation, switch to another agent and start new are “a different line” in experience, but long-term memory still persists across conversations.
• Auto-extract happens after each Q&A round, not when you “end the conversation”.
• Nightly auto-tidy does not depend on you being online, and is unrelated to deleting conversations.

Common questions

Q: Why does the help agent not remember my preferences? A: The help agent has no personal memory. Select a workspace and talk to your own agent, or add entries by hand on that agent’s Memory page.

Q: I said “please remember” and nothing happened? A: Confirm you are on a workspace agent, not the help agent; see whether the assistant confirmed a write; refresh the entry list on Memory. You can also use the message Remember button.

Q: What happens if I say “delete that memory”? A: The assistant should search then delete the matching long-term entry and briefly confirm. You can also tap Delete on that entry on Memory.

Q: Will auto-tidy randomly delete my preferences? A: The policy is conservative (delete cap; stable preferences kept first). Check auto-tidy log for what was deleted; if wrong, create again by hand or correct in chat.

Q: Memory conflicts with a skill or workspace knowledge? A: Knowledge base and skill notes beat generalized memory; on conflict, the clear metric in the knowledge base wins; treat memory as personal preference and a historical snapshot.

Q: Are filenames like memory-xxxx.md normal? A: Yes. Long-term memory bodies live in several files in the system directory; on Memory you see titles and notes. You need not care about disk filenames (implementation).

Related docs

• Product features overview
• Practice guide: where to put your rules — persona / operating rules / memory split
• Practice guide: USER and memory templates
• Look up past conversations — original earlier chat, no need to write memory first
• Agent memory maintenance (admin)
• Messages, conversations, and attachments
• Skills and agents
• Three layers of the knowledge base