OPEN LINK ↗
// 1h agoRESEARCH PAPER
CLAUDE.md Growth Explained by Catastrophic Remembering
A new paper argues that CLAUDE.md and AGENTS.md files grow indefinitely because adding rules is cheap, while deleting them risks regressions when their original rationale is forgotten. It finds that comments preserving rationale can reduce excess instructions by 99.3% and improve agent instruction-following by up to 23.1%.
// ANALYSIS
The paper identifies a painfully familiar maintenance failure: agent instructions behave like code without tests, ownership, or garbage collection. Treating rationale as a first-class artifact could make these files smaller, safer, and easier to prune.
- –Analysis of 247,694 instruction lifetimes across 1,867 repositories found prompts grew 226% over their lifetimes.
- –Older instructions become increasingly unlikely to be deleted, creating a one-way accumulation bias.
- –Comments function like documentation for rules, preserving why an instruction exists and what behavior it protects.
- –The findings suggest teams should review agent instructions like code: test them, document rationale, and remove dead rules.
- –The idea generalizes beyond CLAUDE.md to memory systems, policy files, prompts, and other persistent agent context.
// TAGS
claude-mdagents-mdresearchcontext-engineeringagent-memoryai-codingcoding-agent
DISCOVERED
1h ago
2026-08-13
PUBLISHED
1h ago
2026-08-13
RELEVANCE
9/ 10
AUTHOR
dani_avila7