Agent context tricks: state blocks and anchor reinjection against long-task drift

bigaiguy · x · 2026-09-14

Closing tricks from the agent memory repo: the State Block (before each complex step, the agent writes three lines — known, needed, next — re-grounding itself each turn) and the Anchor Reinjection (re-inject a one-line goal reminder every 3-4 turns; goal-drift cut by more than half in testing).

Related event: Ex-Googler Open-Sources Agent Memory System with 9 Context Tricks(5 posts)→

Original post →

More from coding & agent

coding & agent channel →