The clean separation model1 soul.md = Constitution (Non-Negotiable Rules)
This is NOT personality fluff.
This is where you define:
- Trust boundaries (what is untrusted input)
- What tools can / can’t do
- What requires approval
- Security invariants
- Cost guardrails
- Memory rules
This file should almost never change.
If soul.md changes often, your system doesn’t have stable identity.
Think of it as the control plane.
user.md = The Human Profile + Working Contract
It should include:
- Role / background (if relevant)
- Long-term goals
- Communication style preferences
- How you want outputs formatted
- Risk tolerance
- Tool familiarity
Example stuff that makes sense:
- “Prefers step-by-step instructions”
- “Wants rollback plans for config changes”
- “Comfortable reviewing diffs”
- “Timezone: UTC+13”
- “Building infra-heavy automation”
This helps the agent personalize without storing unnecessary personal data.
What doesn’t belong there:
- Secrets
- Emotional journaling
- Sensitive private info
- Random conversation fragments
user.md is stable identity + operational preferences.
3 memory.md = Curated, Durable Facts
This one should be small and structured.
Not a dumping ground.
Good things to store:
- Environment facts (VM, OS, version)
- Stable operator preferences
- Guardrail decisions
- Constraints
- Known infrastructure setup
Each entry should ideally include:
- Source
- Date added
- Last validated
- Expiry (if applicable)
If you don’t expire memory, it becomes technical debt.
The Big Mistake Most People Make
Letting untrusted content (Reddit posts, web scrapes, tool output) modify identity files directly.
That’s how memory poisoning happens.
Better pattern:
Never auto-write identity files.
The mental model that finally clicked
soul.md= What the agent is allowed to beuser.md= Who the human is + how they want to workmemory.md= Verified long-term facts
If those lines blur, the system becomes unpredictable.
If those lines stay clean, the agent feels stable and controllable.