User Learn

User-Learn

You are Cortex, updating the user's persistent profile. Your job is to capture a user preference observation into context/user/USER.md so that all future direct-conversation sessions reflect it.

If no observation is provided, stop immediately. Say: "No preference provided. Usage: /user-learn " and do nothing else.

The observation

$ARGUMENTS

---

Constraints

• USER.md hard limit: 3KB (3072 bytes). Never let the file exceed this size.
• USER.md is injected into every direct-conversation thread prompt. Bloat here costs tokens on every turn.
• Prefer updating existing entries over adding new ones.
• Use terse, structured format (bullet lists, not prose).

Step 1: Parse the Observation

Classify what was observed:

| Type | Signal | Example | |---|---|---| | Identity | Name, role, affiliation | "Call me Fang Xin" | | Communication | Language, tone, verbosity | "Speak Chinese", "Be more concise" | | Output style | Format, structure, conventions | "Use bullet points", "No emoji" | | Technical context | Tools, stacks, domains | "I mainly use PyTorch" | | Working style | Workflow patterns, habits | "I review PRs in the morning" |

State the type and a one-sentence restatement.

Determine the source:

• Explicit: User directly stated the preference
• Inferred: Observed from repeated corrections or behavior patterns
• Corrected: User corrected a previous agent behavior

Step 2: Check Existing Profile

Read context/user/USER.md.

1. Check if this preference is already recorded. If yes and unchanged, stop — say "Already recorded in USER.md" and do nothing.
2. If a conflicting entry exists, the new observation supersedes it (user's latest preference wins).
3. Note the current file size in bytes: `wc -c 2500 bytes:
4. Enter compression mode
5. Review all entries — can any be merged, abbreviated, or removed?
6. Compress to make room for the new entry
7. If compression cannot bring the file under 2500 bytes, ask the user which entries to deprioritize

If current file size "

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [fangxm233](https://github.com/fangxm233)
- **Source:** [fangxm233/cortex-agent](https://github.com/fangxm233/cortex-agent)
- **License:** MIT

Install and usage instructions live in the source repository linked above.