Agents that run for days lose the rules they were given. We measured how that happens on frontier models from Anthropic and Google, and which remedies bring the work back.
The answer is sharper than gradual decay. Forgetting happens at one moment, when a commitment leaves the context window, and each kind of failure responds to a different remedy. Keeping commitments in force took a regulated task from 43% fidelity to 100%.
What we found
- Distance inside the context made no difference. Sonnet 4.6 read a commitment without error at every distance we tested, up to 12,000 tokens back. Once the commitment left the context window, the model treated it as never stated.
- On a regulated task covering debarment, budget caps, subsidiary structure and eligibility waivers, Nenya scored 1.000. Unaided agents and agents using similarity search both scored 0.429; the best alternative reached 0.886.
- Across 8,192 runs, what recovered a lost rule was largely separate from what taught an agent a rule nobody had stated.
- Picking out binding commitments from text that looked like them was exact at the scale we tested: all 20 real commitments found among 60 decoy turns, with no false alarms.
- The drift shows up in production. In thirty days of the author’s own Claude Code sessions (26,379 records and 36 compactions), commitments drifted where the context was compacted.