Agent memos: file records by tag, look them up when you need them

In user language: for people who get continuous work done with the same work agent. Employee files, leave notes, expense receipts, and business notes you jotted go into memos by tag; when you need them, one more sentence looks them up. Memos are not “memory” recalled automatically across conversations, and they do not come into every round on their own.

For the user-facing product notes, see [help/product-features/agent-memos.md](/docs/help-agent-memos); for the mechanism source of truth, see [docs/core-mechanisms/智能体备忘.md](/docs/mech-agent-memos).

Date: 2026-08-28

The takeaway (read this first)

Across conversations, a work agent automatically brings memory (preferences, confirmed decisions). That set is a poor place to pile “Zhang San’s phone”, “August leave”, “this invoice”. Those records grow, need tags and dates to browse again, and should not occupy context during small talk.

Memos are that small store: text, files, images, filed by tag, with a date attached on save. In chat mode, if you did not clearly say look up or save, the assistant does not browse; after you switch to memo mode, with no special note, “note / look up” targets memos.

In one sentence: memory covers habits that still hold across conversations; memos cover records you take out when you need them.

flowchart TD
  A["About to talk to a work agent"] --> B{"What is this stretch of talk for"}
  B -->|"note an employee / leave / expense / attachment"| C["Write into memos"]
  B -->|"look up Zhang San / look up leave / look up expenses by date"| D["Search memos"]
  B -->|"remember my preference / write into memory"| E["Goes to memory, not memos"]
  B -->|"small talk, ask knowledge, query business numbers"| F["Does not browse memos"]
  C --> G{"Currently in memo mode?"}
  D --> G
  G -->|"yes: no need to say “memo” every time"| H["Read and write memos by default"]
  G -->|"no: must say clearly you want to look up or save memos"| H
  H --> I["The pane refreshes at once; after you select, you can say change or delete"]

Who this is for, and what it solves

It only appears in workspace “My agents”, main-site conversation. The help assistant has no memos. Copying an agent plan also does not copy memos — memos belong to you and this assistant, not a shareable plan.

Typical situations:

• An HR assistant files by person: Zhang San’s phone, department, emergency contact — one record per person; later notes merge into that one;
• Leave and expense are a new event each time — the same tag can have many records, looked up again by save date;
• You drop an invoice or a leave form into the conversation and want the assistant to understand it first, then store it with a description;
• When the count grows, you do not want every round to stuff the whole store into the model; take it only when you open your mouth to look it up.

For people getting work done: say save and it saves, say look up and it looks up; after you select, say change or delete and it acts. For people who only ask “how do I use memos”: they get the steps — memos are not used as background knowledge during small talk.

What memos are and are not

“Tasks” used to be called “task memos”; they are not the same set as memos on this page. Do not write to-dos into memos, and do not write employee files into memory.

How to choose a tag

The tag is the retrieval entry. When you create a tag you pick a type; the same name keeps that type afterwards.

Examples:

• Tag “employee”, unique key “Zhang San” → the whole store has only one Zhang San. Saying “note employee Zhang San: department changed to R&D” merges by default; it does not open another record.
• Tag “leave” → Zhang San takes two days this month, Li Si takes one day: three independent records.
• Tag “expense” → look up by save date “expenses from 1 Aug to 14 Aug”.

Every save automatically writes a save date (the calendar day in your account time zone). If you do not set another title, a unique record often uses “tag · key” (e.g. “employee · Zhang San”).

What you can say to the assistant

Say what you want done, and it acts; ask “how / how do I note memos” and it explains.

In chat mode (the default), if there is no clear intent such as “memo / note employee / look up leave”, the assistant does not browse memos. After you switch to memo mode, you need not say “memo” every time; when you want another path, say clearly “write into memory” or “write into tasks”.

Two conversation modes

Chat mode (default)

Everyday work uses this. Memory, knowledge, and skills work as usual. The assistant only touches memos when you clearly want to save or look them up. Fits “occasionally note one, occasionally look one up”, without letting the store interrupt ordinary Q&A.

Memo mode

In the composer, switch the mode to Memo. After that:

1. A memo pane appears on the left of the conversation, listing this assistant’s existing records; when the assistant changes them, it refreshes at once.
2. “Note / look up / check / list / delete” target memos by default.
3. After you close the pane, a leftover Memo button remains and can be dragged elsewhere; tap it, or tap the memo-pane icon in the composer, to open it again — no need to switch mode a second time.
4. The pane can be dragged, and you can drag the edge to resize; the next open keeps the last position and size.
5. In the checkboxes on the left of the pane, select one or more records (up to eight); a chosen tag appears in the composer. Say change or delete directly. Tap × on the tag to unselect.
6. The reading icon on the right of each record opens a larger window; Markdown is laid out; long text can go fullscreen.

In memo mode you can still say clearly “write into memory”. For small talk, changing persona, or querying business numbers, the assistant still does not browse memos.

How to see and manage them in the UI

1. Say it in Messages (the main path): select your own work agent, and speak as in the previous section. With a file or image, attach it first, then “save into memos”.
2. Memo pane: appears after you switch to memo mode. Fits checking the record you just wrote while you talk; after you select, change and delete are fastest.
3. Agent detail → Memos: browse all records, filter by tag or date, add or delete text memos by hand, manage tags. Opening a record defaults to laid-out reading; tap edit to change; fullscreen is available. Files and images are still better given to the assistant in conversation, so it can write a content description before storing.

You must select a workspace first, then your own agent. The help assistant’s conversation has no Memos page, and cannot switch to memo mode.

How files and images enter memos

Plain text can be “note…” directly. With an attachment:

1. Attach the file or image in this round;
2. Say “save into memos” (in memo mode, “note” is enough);
3. The assistant understands the content first (images from the picture, files by reading), then writes a content description together with the original.

The description is for later retrieval: looking up “high-speed rail ticket” relies on the points written at the time, without parsing the original again every round. When covering an existing unique record, update the description the same way.

Content can be plain text, a file, an image, or mixed.

Where it lands (for people aligning the implementation)

Memos do not go through MEMORY.md in the runtime folder. Memory is Markdown files; memos are tables in the workspace shard database:

A unique tag has only one record on “same assistant + tag name + unique key”. Retrieval can use tag, unique key, save-date range, and keywords in title/body/description.

The conversation request carries whether memo mode is on, and the record ids selected in the pane; this round’s change and delete target those ids, without searching first. After the assistant changes memos, the UI sends a memos-changed notice, and the pane reloads at once.

Copying a plan, the help assistant, and visitor paths in the site widget do not carry this store. Finer conventions are in the mechanism article.

When not to force memos

• Preferences you want every future conversation to follow — say “please remember”; do not note them as memos.
• Policy and metric the whole team should read — write into the knowledge base (workspace layer for every assistant in the workspace, user layer for personal material).
• To-dos and goal drafts — use tasks; do not mix them with leave and expense under one tag.
• Asking the help assistant how to use the product — that assistant has no memos; to file records, select your work agent first.
• Only what this one conversation is doing — write it in the topic; the conclusion after it is done goes in review.

Purpose goes in the topic, the conclusion after the work is done goes in review, habits across conversations go in memory, records you look up when needed go in memos, team metric goes in the knowledge base.

Related reading

• Agent memos — tags, modes, pane selection, and implementation map
• Product features: agent memos — usage notes in help
• Agent memory — why “remember” and memos are not the same thing
• Where work-agent memory is stored — memory lands in a folder, memos land in database tables
• Conversation topics — the topic covers “what we are doing now”
• Conversation review — how to leave a conclusion after the work is done
• Product introduction — where Messages and work agents sit in the product