Research2026-07-01
Researchers described Self-GC, a method for deciding what a long-running AI agent should keep in its working memory. Instead of deleting the oldest text or writing a summary at the limit, it treats past turns, tool results and files as tracked items that a separate planner can fold, hide or discard, with the option to restore them. Reported tests covered 33 difficult sessions and 332 sessions drawn from production, plus a live deployment.
What changed
Agent systems trimmed old context with simple rules such as dropping the oldest messages, or wrote a summary near the limit.
What it unlocks
Managing an agent's accumulated working memory as tracked, recoverable items rather than deleting text after the fact.
- 43.95% of prefix tokens pruned
- 91.27% to 94.58% no-impact rate
- 10% to 15% fewer daytime input tokens
What you need to act on it
- engineering work to implement the described method
Sources