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).
More from coding & agent
- Gemini 4 tops public coding benchmarks but drops to bottom on private tasks — Hesamation · 2026-10-30
- User builds 20-bot Grok suite to track wins, prep 1:1s and close promises — NicoVerderosa · 2026-10-30
- Why 30-year-old BM25 beats vector embeddings for agentic search, per search vet Jo Kristian Bergum — jobergum · 2026-10-30
- LangChain CEO opens Interrupt London with keynote on why orgs should own their intelligence — LangChain · 2026-10-30
- Simlock: open-source control plane gives parallel coding agents their own iOS/Android simulators — Scobleizer · 2026-10-30
- Keyfleet launches open canvas where personal AI agents team up with a shared USDC treasury — seanwbren · 2026-10-30