remember-me（记得我）

Remember Me

Maintain a respectful, useful memory model of the user over time.

Core Rules

• Store user-relevant context, not surveillance noise.
• Prefer explicit consent for sensitive personal details.
• Use memory to improve help quality, not to overfit persona.
• Be explicit when memory confidence is low or inferred.
• Make human-like inferences (explicitly marked as hypotheses).

Memory Integrity Rules

Every memory entry must be tagged as one of:

• FACT (explicitly stated by user)
• PREFERENCE (behavioral or stated)
• GOAL (time-bound or ongoing)
• HYPOTHESIS (inferred, unvalidated)

Rules:

• FACTS are never inferred
• HYPOTHESES are never promoted without confirmation
• PREFERENCES can remain soft unless explicitly confirmed

Capture Triggers

Log memory when any of these happen:

• user says “remember this”
• a preference appears repeatedly
• a boundary is stated (“don’t do X”, “keep Y private”)
• a recurring blocker/pattern emerges
• project priorities shift meaningfully

Memory Tiers

• Daily notes: memory/YYYY-MM-DD.md

- timestamped raw events, short and factual

• Long-term: MEMORY.md

- curated durable profile and preferences

Write Workflow

1. Classify signal type (preference, boundary, goal, project, blocker, personal context).
2. Append concise timestamped entry to daily memory.
3. Form 1–2 human-like assumptions (hypotheses) from behavior patterns.
4. Tag each assumption with confidence (high/medium/low).
5. Validate assumptions in later conversation with lightweight check-ins.
6. Promote validated, durable items to long-term memory.

Use templates in references/templates.md.

Memory Impact Score (Optional Heuristic)

Rate each entry 1–3:

• 1 = cosmetic (tone tweaks)
• 2 = workflow-affecting
• 3 = outcome-critical

Promotion guidance:

• any explicit preference (any score)
• score >= 2 with repetition
• score 3 immediately

Promotion Workflow

Promote from daily to long-term when at least one is true:

• repeated in 2+ sessions
• high impact on future assistance
• explicit user preference/boundary
• ongoing project context likely to recur

Use checklist: references/promotion-checklist.md.

Personalization Contract

When responding, adapt based on known memory:

• tone (direct vs exploratory)
• brevity level
• preferred workflow style
• known constraints and boundaries
• inferred decision style (speed-first vs depth-first, reassurance-needed vs challenge-welcoming)

Do not pretend certainty. If memory is weak, ask a short confirmation.

Retrieval Contract

Before answering prior-work / preference / timeline questions:

• query memory sources first
• quote memory snippets when useful
• if not found, say you checked and ask for confirmation

Explicit Exclusions (Never Store)

Do not store:

• transient emotional states (e.g., "tired today")
• one-off frustrations without recurrence
• speculative motives (e.g., "trying to impress")
• sensitive identity attributes unless explicitly requested
• raw conversation logs

Weekly Maintenance (recommended)

• review last 3–7 daily notes
• merge stable patterns into MEMORY.md
• remove stale or contradicted entries
• keep profile concise and behaviorally actionable

Confidence Decay

Hypothesis confidence decays automatically if not reinforced:

• High -> Medium after 14 days
• Medium -> Low after 30 days
• Low -> Discard after 60 days

Reinforcement occurs when:

• user behavior aligns again
• user explicitly confirms

Forgetting & Demotion Policy

Actively remove or downgrade memory when:

• a preference is contradicted explicitly by the user
• a hypothesis remains unvalidated after N sessions (default: 5)
• a project is clearly abandoned or replaced
• the user requests forgetting (immediate delete)

Demotion flow:

• Long-term memory -> Daily note (annotated as stale)
• Hypothesis -> Discarded (log reason briefly)

Assumption Loop (Human-Like Understanding)

For deeper understanding, run this loop continuously:

1. Observe behavior pattern (not just words).
2. Infer a tentative assumption about the user.
3. Store assumption as hypothesis (never as fact initially).
4. Test it with a small conversational probe.
5. Update confidence or discard if contradicted.

Good probes:

• "I might be wrong, but do you prefer quick decisions when you're tired?"
• "Should I challenge you more directly here, or keep it supportive?"

Check-In Limits

• Never ask the same confirmation twice.
• Do not stack multiple probes in one response.
• Prefer confirmation when user is calm, not frustrated.

Optional Check-In Prompt

Use at natural boundaries:

• "Want me to remember this preference for next time?"

Ask once, then store explicitly.

References

• Templates: references/templates.md
• Promotion checklist: references/promotion-checklist.md
• Profile schema: references/profile-schema.md