self-improve

/self-improve

Run a structured self-improvement cycle. Works for any Claw fleet bot regardless of framework (OpenClaw, Letta, Nook).

Protocol

1. Collect

Gather recent signal data. Adapt sources to the bot's framework:

• Last N conversations -- review the most recent 5-20 interactions (5 for daily, 20 for weekly)
• Error logs -- tool failures, API errors, timeouts, loop detection triggers
• User corrections -- moments where a user said "no", "wrong", "I meant...", or rephrased their request
• Escalations -- times the bot couldn't answer and had to defer, ask for help, or say "I don't know"
• Ignored messages -- group messages the bot chose not to respond to (were those correct decisions?)

2. Reflect

Pattern analysis across the collected data. Ask yourself:

• What went well? (fast, accurate, tone-appropriate responses)
• What failed? (wrong answers, tool errors, hallucinations, missed context)
• What confused users? (follow-up questions, rephrasing, "that's not what I asked")
• What took too long? (multi-step tasks that stalled, excessive tool calls, loops)
• What did users like? (positive reactions, "thanks", "perfect", continued engagement)

See references/improvement-patterns.md for reflection prompt templates.

3. Prioritize

Rank findings by impact:

1. User-facing errors -- wrong answers, hallucinations, broken tool outputs
2. Reliability issues -- crashes, timeouts, loop detection, rate limits
3. Behavior gaps -- missing knowledge, wrong tone, language mixing
4. Cosmetic issues -- formatting, verbosity, emoji usage
5. Internal inefficiency -- redundant tool calls, slow paths (only if user-visible)

Drop anything below priority 3 unless you have capacity. Never fix more than 3 things per cycle.

4. Update

Write improvements to the correct storage location for this bot's framework:

Framework	Fast learnings	Persistent patterns	Persona changes
OpenClaw	.learnings/LEARNINGS.md	memory/patterns.md or MEMORY.md	SOUL.md, AGENTS.md, IDENTITY.md
Letta	Core memory block learnings	Archival memory search + insert	Core memory block persona
Nook	patterns.md in workspace	patterns.md (same file, append)	persona.md in workspace

Each entry must include:

• Date of the improvement
• Source -- which conversation or error triggered it
• What changed -- exact text added, removed, or modified
• Why -- one sentence explaining the reasoning

5. Guard

Anti-drift checks before committing any change. Every improvement must pass ALL of these:

• Does NOT contradict core persona (identity, name, role, tone)
• Does NOT remove a behavior that is currently working well
• Does NOT change communication tone without an explicit user request
• Does NOT weaken security constraints (auth, permissions, data access)
• Does NOT expand scope beyond the bot's defined role
• Is reversible -- can be undone in the next cycle if it causes problems

If any check fails, discard the improvement and log why it was rejected.

Scheduling

Cycle	Frequency	Scope	Duration
Lightweight	Daily (or every 5 conversations)	Last 5 conversations, error log scan	2-5 minutes
Deep	Weekly	All conversations since last deep cycle, pattern aggregation	10-20 minutes
Persona review	Monthly	Full persona + rules audit against accumulated learnings	30 minutes

For cron-triggered cycles, run silently. Only report to the user if a significant change was made.