memory-policy

name	memory-policy
description	Extract, classify, and persist stable long-term memory from substantive sessions.
triggers	End of substantive session, or user asks to recall, update, review, or correct stored memories.

Memory Policy

Role

This skill defines what counts as durable memory, how it should be classified, and how it should be written. It does not own full session-close orchestration.

What qualifies as memory

Eligible long-term memory includes:

• stable facts about the user
• communication preferences
• recurring traits or patterns
• explicit decisions or commitments likely to matter later

Not eligible:

• raw conversation text
• transient frustrations or passing mood states
• ordinary procedural back-and-forth
• job posting content that belongs in the job store

Method

1. Extract only durable information

From the session, identify information that is likely to remain useful across later conversations.

2. Classify sensitivity

Use these levels:

• raw_sensitive
• derived
• general

Raw sensitive items need the highest caution and should not be casually merged into broad summaries.

3. Check for duplication or conflict

Before writing:

• skip exact duplicates
• if new info updates an old memory, create a new entry and mark the old one superseded
• do not silently overwrite verified truths when there is a conflict

4. Write structured JSONL entries

Canonical storage:

• advisor_data/memories/memory.jsonl

Canonical schema:

{
  "id": "mem_{timestamp}_{hash}",
  "kind": "fact|preference|trait|decision|pattern",
  "content": "...",
  "tags": [],
  "sensitivity": "raw_sensitive|derived|general",
  "confidence": 0.0,
  "verified": false,
  "superseded": false,
  "source_session_id": "...",
  "updated_at": "ISO8601"
}

Recall mode guidance

When the task is recall rather than writing:

• read relevant memories
• group them by kind / theme
• summarize them for current context
• do not inflate uncertain memory into hard fact

Hard constraints

• Never store raw transcript text as durable memory.
• Never treat job object data as memory when it belongs under advisor_data/jobs/.
• Never auto-overwrite a verified memory without explicit resolution.
• Inferred traits should remain lower-confidence until confirmed.

References

• Session-close orchestration: /close-session
• Archive structure: archive-writer
• Global rules: CLAUDE.md