Practice guide: USER.md and memory — templates and split
user profile (USER.md), long-term memory, and daily notes are all “information about you”, but how you maintain them and how chat injects them differ. All live on Agents → Memory (
Source help/en/guides/user-and-memory-templates.md
In user language: user profile (USER.md), long-term memory, and daily notes are all “information about you”, but how you maintain them and how chat injects them differ. All live on Agents → Memory (profile at the top, entries below).
How the three split
| User profile USER | Long-term memory | Daily notes | |
|---|---|---|---|
| Typical content | Who I am, how to address me, long-term background | Preference corrections, decisions, milestones, where to look | What I am doing today / these days |
| Maintain | Hand-edit a Markdown block | Form entries, Remember, auto-extract | Remember pick daily note, compress archive |
| Structured? | One free-text block | Title + type + notes + body | One article per day |
| Example | “I am finance; reports only in Excel” | “Later export CSV, not PDF” | “This afternoon I am changing the login page” |
Operating rules hold steps that must run every time; do not replace AGENTS with USER or memory.
USER.md templates
Template A: workplace identity
# About me
- Address: Xiao Li (Li Gong is also fine)
- Role: backend, order service; frontend only basics
- Communication: technical detail can go a bit deep, but conclusion first
- Do not need: long marketing copy, unrelated industry gossip
Template B: business role
# About me
- Department: East China sales ops
- Systems I use often: CRM, internal BI (concrete query steps are in workspace skills; do not repeat here)
- Taboo: do not disclose unpublished price policy to external customers
Keep the profile short and stable; changing project milestones belong in memory · project progress.
Long-term memory entry templates
When writing, pick a type on Memory, or pick Long-term memory when tapping Remember on a message.
Preferences and corrections (feedback)
Title example: Answer format: conclusion then list
Body:
**Why:** Long paragraphs are hard to scan quickly.
**How to apply:** Prefer bullets except in code blocks; use numbered lists when more than 5 steps.
About me (user) — vs USER
| Put in USER | Put in long-term memory · About me |
|---|---|
| One overview block, rarely changed | Separate entries, searchable, deletable one by one |
| Address, department, long-term role | “I have a Go background, Rust beginner” as its own entry |
They can coexist; avoid exact duplicates.
Project progress (project)
Title: Q2 launch milestone
Body:
- Deadline: 2026-06-30 (absolute date)
- Scope: Web only, not mobile
- Status: testing (updated 2026-06-15)
Where to look (reference)
Title: API spec entry
Body:
Team wiki “Open platform / REST conventions” chapter; concrete paths follow the workspace knowledge base — do not store URL detail here.
Daily-note example
## 2026-06-15
- Morning: aligned login-redesign copy with product
- To confirm: whether error copy unifies to “please check account or password”
- Next conversation: continue SSO option comparison
What is MEMORY.md?
What you see in the UI is memory entries; in the background the system maintains a long-term memory list (implementation: MEMORY.md index + bodies under memory/entries/). You need not hand-edit the index; add/change/delete on Memory.
Corrections and cleanup
| Approach | When |
|---|---|
| In chat say “that was wrong / forget / delete that memory” | The assistant searches then updates or deletes the matching long-term memory |
| Edit / delete an entry on Memory | You want precise control of one entry |
| Tidy my memory / nightly auto-tidy | Drop obvious duplicates and expired daily notes; recorded in auto-tidy log |
Details: Agent memory.
Related docs
- Agent memory — Remember, correct, auto-tidy, injection rules
- Where to put your rules
- File map