Skip to main content
Ultron doesn’t inject all memories into every conversation. That would flood context with irrelevant information. Instead, a targeted recall system selects exactly the entries most relevant to the current task.

​
How recall works

1

Side-query runs

Before each agent turn, a separate fast-model side-query runs against the memory store. It receives the current task objective and searches for relevant entries.
2

Candidates scored

Memory entries are ranked by semantic relevance to the current objective. Recency is a secondary factor — fresh entries score slightly higher when relevance is equal.
3

Deduplication applied

Entries that were already surfaced in recent turns are skipped. This prevents the same memory from appearing repeatedly and wasting context.
4

Top 5 injected

The 5 highest-scoring entries are injected into context before Kimi K2 sees the message. More than 5 adds noise without improving output.

​
Why the cap is 5

Intuition says more context is better. In practice, it isn’t. With more than 5 memories injected, two things happen:
  1. Irrelevant entries dilute the signal of relevant ones
  2. Token budget fills up faster, triggering compression sooner
5 high-signal entries outperform 15 mixed-relevance entries every time.

​
Memory drift defense

Before any memory entry is injected, Ultron validates it:
  • File paths mentioned — do they still exist?
  • Functions referenced — are they still in the codebase?
  • External references — are they still current?
Stale memories that reference deleted files or outdated context are flagged or skipped rather than injected and acted on incorrectly.
If Ultron says “I don’t have recent information on this” when you expect it to, the relevant memory may be stale or tagged with a different topic. Use the Brain Graph to audit what’s stored.