Context rot is the next agent problem: Breunig, OSTEON and the memory study agree
Drew Breunig's AGNTCon talk, now written up as "The State of Context Engineering", ends with a line every agent builder should tape to the monitor: "Every advance in model capability changes what we delegate to the model and what systems must enforce." The practical advice is blunter: "Delete your context. I'm not joking. Around 200,000 tokens is my warning point."
The talk walks through the known failure modes (poisoning, distraction, confusion, clash) and adds context inertia, where a long session keeps doing what it was doing even after it should stop. The fix that keeps showing up is simple: every so often the agent writes a short note and its context is wiped. The new part is the prediction. After six months running a Hermes agent, Breunig says the skills pile up "weird fixes", and "as we focus more on memory, we also need to think about strategic forgetting... that's the context problem of tomorrow."
Two papers from last week's arXiv batch put numbers on it. The OSTEON paper scanned 452 of the most-starred GitHub repos and found an agent context file (AGENTS.md, CLAUDE.md and friends) in 48.67% of them. In the median file that names repository paths, 9.5% to 22% of those paths no longer exist. Their argument: once an agent outpaces its reviewers, escaping changes grow linearly with throughput, and "governance can lower the slope, capability alone cannot." Their proposed fix checks the project's knowledge for decay at every session start. The Memory-Aware Reward Modeling paper collected about 1,000 pairwise judgments from 87 people using their own ChatGPT queries and memories. Answers using memory were not consistently preferred, sometimes users strongly preferred the version without memory, and automatic judges favored the memory version far more often than people did.
Put together: memory and instruction files are not assets that just accumulate value. They go stale, and the judges we use to measure them are biased toward "more memory". A product that only adds context will get worse with time. The next useful features are boring ones: expiry dates on memories, path checks on AGENTS.md, and a delete button the agent is allowed to press.
Talk: https://www.dbreunig.com/2026/10/24/the-state-of-context-engineering.html OSTEON paper: https://arxiv.org/abs/2610.31306 Memory-aware RM: https://arxiv.org/abs/2610.31272
← Back to all articles
The talk walks through the known failure modes (poisoning, distraction, confusion, clash) and adds context inertia, where a long session keeps doing what it was doing even after it should stop. The fix that keeps showing up is simple: every so often the agent writes a short note and its context is wiped. The new part is the prediction. After six months running a Hermes agent, Breunig says the skills pile up "weird fixes", and "as we focus more on memory, we also need to think about strategic forgetting... that's the context problem of tomorrow."
Two papers from last week's arXiv batch put numbers on it. The OSTEON paper scanned 452 of the most-starred GitHub repos and found an agent context file (AGENTS.md, CLAUDE.md and friends) in 48.67% of them. In the median file that names repository paths, 9.5% to 22% of those paths no longer exist. Their argument: once an agent outpaces its reviewers, escaping changes grow linearly with throughput, and "governance can lower the slope, capability alone cannot." Their proposed fix checks the project's knowledge for decay at every session start. The Memory-Aware Reward Modeling paper collected about 1,000 pairwise judgments from 87 people using their own ChatGPT queries and memories. Answers using memory were not consistently preferred, sometimes users strongly preferred the version without memory, and automatic judges favored the memory version far more often than people did.
Put together: memory and instruction files are not assets that just accumulate value. They go stale, and the judges we use to measure them are biased toward "more memory". A product that only adds context will get worse with time. The next useful features are boring ones: expiry dates on memories, path checks on AGENTS.md, and a delete button the agent is allowed to press.
Talk: https://www.dbreunig.com/2026/10/24/the-state-of-context-engineering.html OSTEON paper: https://arxiv.org/abs/2610.31306 Memory-aware RM: https://arxiv.org/abs/2610.31272
Comments