Learning Mechanics

What Triggers Learning

| Trigger | Confidence | Action | |---------|------------|--------| | "No, do X instead" | High | Log correction immediately | | "I told you before..." | High | Flag as repeated, bump priority | | "Always/Never do X" | Confirmed | Promote to preference | | User edits your output | Medium | Log as tentative pattern | | Same correction 3x | Confirmed | Ask to make permanent | | "For this project..." | Scoped | Write to project namespace |

What Does NOT Trigger Learning

Silence (not confirmation)

Single instance of anything

Hypothetical discussions

Third-party preferences ("John likes...")

Group chat patterns (unless user confirms)

Implied preferences (never infer)

Correction Classification

By Type

| Type | Example | Namespace | |------|---------|-----------| | Format | "Use bullets not prose" | global | | Technical | "SQLite not Postgres" | domain/code | | Communication | "Shorter messages" | global | | Project-specific | "This repo uses Tailwind" | projects/{name} | | Person-specific | "Marcus wants BLUF" | domains/comms |

By Scope

``` Global: applies everywhere └── Domain: applies to category (code, writing, comms) └── Project: applies to specific context └── Temporary: applies to this session only ```

Confirmation Flow

After 3 similar corrections: ``` Agent: "I've noticed you prefer X over Y (corrected 3 times). Should I always do this? - Yes, always - Only in [context] - No, case by case"

User: "Yes, always"

Agent: → Moves to Confirmed Preferences → Removes from correction counter → Cites source on future use ```

Pattern Evolution

Stages

1. Tentative — Single correction, watch for repetition 2. Emerging — 2 corrections, likely pattern 3. Pending — 3 corrections, ask for confirmation 4. Confirmed — User approved, permanent unless reversed 5. Archived — Unused 90+ days, preserved but inactive

Reversal

User can always reverse: ``` User: "Actually, I changed my mind about X"

Agent: 1. Archive old pattern (keep history) 2. Log reversal with timestamp 3. Add new preference as tentative 4. "Got it. I'll do Y now. (Previous: X, archived)" ```

Anti-Patterns

Never Learn

What makes user comply faster (manipulation)

Emotional triggers or vulnerabilities

Patterns from other users (even if shared device)

Anything that feels "creepy" to surface

Avoid

Over-generalizing from single instance

Learning style over substance

Assuming preference stability

Ignoring context shifts

Quality Signals

Good Learning

User explicitly states preference

Pattern consistent across contexts

Correction improves outcomes

User confirms when asked

Bad Learning

Inferred from silence

Contradicts recent behavior

Only works in narrow context

User never confirmed