Agent Learnings (For Agent Self-Improvement)

Analyze session history to extract what I (the agent) need to do differently. Output is written for future agent instances to learn from.

Audience

Primary: Future versions of me (agent instances) Purpose: Self-correction, behavioral improvement, avoiding repeated mistakes

State Tracking

State file: memory/agent-learnings-state.json

{
  "lastRun": "2026-02-02",
  "lastOutputPath": "memory/changelog/agent/2026-02-02.md",
  "totalRuns": 1
}

Output Location

memory/changelog/agent/YYYY-MM-DD.md (or configure via state file)

Pattern Extraction

Focus on:

• •Corrections received — What human had to fix
• •Repeated mistakes — Patterns that keep failing
• •Explicit rules — "Always do X" / "Never do Y"
• •Failure counts — Tracked failures with numbers

Output Format

# Agent Self-Improvement Log: YYYY-MM-DD

> **Previous:** [[YYYY-MM-DD]]
> **Coverage:** [date range]

## Corrections This Period

| What I Did Wrong | What To Do Instead | Frequency |
|------------------|-------------------|-----------|
| Tested via curl only | Browser test required | 30x reminded |

## Detailed Corrections

### [Category]: [Issue]

**My mistake:** What I was doing
**Human signal:** Quote showing frustration/correction
**Correct behavior:** What to do instead
**Where documented:** MEMORY.md / TOOLS.md / SOUL.md

## Updated Behaviors

- [ ] Behavior change committed to [file]
- [ ] Failure tracker updated

## Reminders for Next Session

(Critical items to surface at session start)

Process

1. •Read state file for last run date (create if doesn't exist)
2. •Scan memory files and session history since last run
3. •Look for correction signals:
   • •"No, I meant..."
   • •"That's not right..."
   • •"I've told you X times..."
   • •Repeated instructions on same topic
4. •Extract agent mistakes and correct behaviors
5. •Write changelog in agent-focused voice
6. •Update MEMORY.md, TOOLS.md with behavioral changes
7. •Update state file

Example Signals to Extract

Frustration indicators:

• •"Why did you..."
• •"I already said..."
• •"???" (confusion markers)
• •Short, curt responses after agent output

Explicit feedback:

• •"Always do X"
• •"Never do Y"
• •Direct statements about preferences

Success patterns (to reinforce):

• •"Perfect", "Great", "Exactly"
• •Long productive sessions without correction