Executive summary
Agents that run for days lose the rules they were given. Experimental Foundations measures how that happens on frontier models from Anthropic and Google, and which remedies recover the work. The experiments find that forgetting happens at one moment, when a commitment leaves the context window, and that each kind of failure responds to a different remedy. Holding commitments in force took a regulated task from 43% fidelity to 100%. The results pinpoint where an agent needs support, and which support it needs.
Findings
- Distance inside the context doesn’t matter. Sonnet 4.6 read a commitment without error at every distance tested, up to 12,000 tokens back, then treated it as never stated once it left the context window.
- 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, and the best alternative reached 0.886.
- Across 8,192 runs, what recovered a lost rule was largely separate from what taught the agent a rule nobody had stated.
- Picking out binding commitments from look-alike text was exact at the tested scale: all 20 real commitments found among 60 decoy turns, with no false alarms.
- The drift is real in production. Across thirty days of the author’s own Claude Code sessions (26,379 records and 36 compactions), commitments drifted where the context was compacted.
These measurements are why Nenya keeps a commitment in force at exactly the moment the context window would drop it.