A preprint analyzing 1,867 public GitHub repositories found that agent context files like CLAUDE.md and AGENTS.md grow on average by 226% over their lifetime, tracked across 247,694 instruction instances. Median files end with 39 instructions, while the 90th percentile reaches 131. Researcher Kushal Chakrabarti calls the pattern 'catastrophic remembering': teams add rules after failures but later lose the original rationale, making files bloated and risky to prune. For builders, this uncontrolled growth directly impacts token usage, latency, and maintenance costs. While the paper's controlled experiment tests a split between executor instructions and maintainer rationale, the repository history cannot confirm whether rationale actually disappears. The implication is that context files are becoming legacy code that needs deliberate curation, versioning, and deduplication tooling to stay manageable.
A preprint by Kushal Chakrabarti tracked 247,694 instruction lifetimes across 1,867 public GitHub repositories. The mean instruction count in CLAUDE.md, AGENTS.md and copilot-instructions.md rose by up to 226% over a file's lifetime, with median files ending at 39 instructions and the 90th percentile at 131.
Key takeaway
Context files are becoming legacy code; their growth is structural and likely to continue, requiring deliberate curation strategies.
What happened
A preprint by Kushal Chakrabarti, reported by Towards AI, tracked 247,694 instruction lifetimes across 1,867 public GitHub repositories. The study found that context files like CLAUDE.md, AGENTS.md, and copilot-instructions.md experience a mean instruction count increase of 226% over their normalized lifetime.
Median files ended with 39 instructions, while the 90th percentile contained 131. Chakrabarti calls this pattern 'catastrophic remembering,' where maintainers add rules after failures but later lose the original incident, failed attempt, or rationale, making files harder to prune safely.
Evidence
Study tracked 247,694 instruction lifetimes across 1,867 public GitHub repositories.
Towards AI · attributed
A preprint by Kushal Chakrabarti tracked 247,694 instruction lifetimes across 1,867 public GitHub repositories.
Mean instruction count rose by up to 226% over a file's lifetime.
Towards AI · attributed
The mean instruction count in CLAUDE.md, AGENTS.md and copilot-instructions.md rose by up to 226% over a file's lifetime, with median files ending at 39 instructions and the 90th percentile at 131.
Researcher Kushal Chakrabarti calls the growth pattern 'catastrophic remembering'.
Towards AI · attributed
Kushal Chakrabarti calls the pattern “catastrophic remembering” in a single-author preprint posted Aug. 11, 2026.
The controlled experiment tested a split that separates executor instructions from maintainer rationale.
Towards AI · attributed
Its controlled experiment tested a split that agent context files rarely offer: instructions for the executor, rationale for the next maintainer.
Why it matters
Builders must anticipate that context files will become major contributors to token usage and latency, necessitating new tooling for versioning, deduplication, and lifecycle management of agent instructions.
Limits and uncertainties
The repository histories do not record whether rationale actually disappeared; the assigned experiment tests the mechanism in benchmark worlds.
The study is a single-author preprint, so it has not undergone peer review.
Practical implications
Operators should expect context files to become significant contributors to token usage and latency, prompting investment in tooling for versioning, deduplication, and lifecycle management.
Teams should establish explicit policies for when to prune or archive rules, relying on the proposed split between executor instructions and maintainer rationale to preserve context.
What to watch
Monitor adoption of the proposed split (executor vs. maintainer instructions) in open-source agent frameworks.
Watch for tooling that automatically tracks and removes obsolete instructions based on failure and evidence history.
