Every error an Ultron agent encounters becomes a lesson. Every lesson is recalled before the next run. Errors happen at most once.
How it works
Error occurs
An agent encounters a failure: an API rate limit, a scraping block, a tool returning an unexpected format, a cold email failing the quality gate.
Lesson stored
The agent immediately stores a feedback memory:{
"type": "feedback",
"tags": ["lesson", "specter"],
"text": "LESSON: Apollo returns 429 on bulk enrichment over 10 leads. PREVENTION: Batch max 10 leads per run, wait 2s between batches."
}
Next run starts
Before starting any work, the agent searches memory for lessons tagged with its name:search_memories({ query: "lesson specter", limit: 10 })
Lessons applied
The retrieved lessons are injected before any tool calls. The agent adjusts its behavior — batching correctly, using alternative approaches, avoiding known failure patterns.
Real examples
| Error | Lesson stored |
|---|
| Apollo 429 on bulk enrichment | Batch max 10, wait 2s between batches |
| ZeroBounce returns catch-all for a domain | Flag domain as catch-all, reduce outreach priority |
| LinkedIn scrape blocked | Switch to Apollo for profile data on this domain |
| Cold email failed quality gate 3x | Default to 2-sentence body for technical audiences |
| Stripe API key expired mid-session | Check Stripe token validity at session start |
Why both errors AND successes
The lessons system stores confirmed patterns too:
VALIDATION: The 3-sentence body + open question CTA
outperformed longer formats. Reply rate: 12% vs 4%.
Keep as default for SaaS founder ICP.
You can manually add lessons. In chat, say: “Remember that [thing you learned] — apply this going forward.” Ultron stores it as a feedback memory immediately.
