What gets stored
Four memory types
| Type | Stores | Example |
|---|---|---|
| user | Your role, background, preferences | ”I’m a B2B SaaS founder, ICP is Series A CTOs” |
| feedback | Corrections AND confirmations | ”Don’t use em-dashes” / “That cold email structure works” |
| project | Non-code business context | ”Q2 freeze starts April 15, no new features after that” |
| reference | Research outputs, external pointers | Competitive analysis on HubSpot, meeting prep for Stripe |
Feedback memory stores both errors and validations. When you confirm something works, Ultron stores that too. This prevents behavior drift — Ultron doesn’t unlearn what’s working.
How memory is used
At the start of every agent turn, a separate Sonnet-powered side-query runs against the memory store. It selects the 5 most relevant entries for the current task. Those entries are injected into context before Kimi K2 sees the message. You don’t manage this. It happens automatically.Memory Layers
The four memory types and when each is created.
Brain Graph
Visual map of everything Ultron knows about your business.
Memory Recall
How Ultron selects the right memories per turn.
Compression Engine
Three-layer cascade that keeps sessions running indefinitely.