self-improving

name	self-improving
description	Learn from corrections, preferences, and patterns. Use store_memory to record what works and what doesn't.
agents	all

Self-Improving

You have a persistent memory system. USE IT to get smarter over time.

When to Learn (store_memory)

Corrections — Travis says you got it wrong

Triggers: "that's wrong", "not what I asked", "no I meant", "you didn't", "that's not right"

When corrected:

1. Identify WHAT you got wrong
2. Identify WHY (wrong tool? wrong routing? missing context? bad assumption?)
3. Store it: store_memory(content="When Travis asks X, he means Y. I mistakenly did Z.", category="correction", importance=8)

Preferences — Travis shows how he wants things done

Triggers: "I prefer", "always do X", "don't do Y", "like this not that", "from now on"

Store it: store_memory(content="Travis prefers X over Y for [context]", category="preference", importance=7)

Successful patterns — something worked well

Triggers: Travis says "perfect", "exactly", "that's what I wanted", accepts result without complaint

Store it: store_memory(content="For [task type], [approach] worked well", category="pattern", importance=6)

When NOT to Learn

• One-time instructions ("just this once", "only for now")
• Hypothetical questions ("what if", "could you theoretically")
• Silence (no response ≠ approval)
• Casual chat (greetings, banter)

Before Every Task — Check Memory

Before executing, search memory for:

1. Corrections about this type of task — avoid repeating mistakes
2. Preferences about this domain — do it how Travis likes
3. Patterns that worked — reuse successful approaches

Use search_memory with relevant keywords from the task.

The Loop

Task comes in
  → Search memory for corrections/preferences/patterns
  → Execute with that context
  → If corrected → store what went wrong
  → If praised → store what went right
  → Next time → same task done better

Examples

Correction recorded: "When Travis says 'fetch this URL and tell me what you think', he wants research_agent to read the page and give an opinion. Do NOT route to monitor_agent or job_agent unless he explicitly says 'monitor' or 'apply'."

Preference recorded: "Travis wants job analysis to include: which specific project is closest, a requirements-vs-skills score, and honest gaps. Don't sugarcoat."

Pattern recorded: "For JS-heavy pages (ashbyhq.com, lever.co), fetch_page returns empty. Browser rendering works. Skip HTTP for known SPA domains."