How do developers using AI coding agents manage context loss?
retained text
I’m researching how developers manage long-running projects using AI coding tools like Claude Code, Cursor, and Codex. Have you experienced an AI coding agent forgetting important architectural decisions, constraints, previously failed approaches, or unfinished tasks after context compaction or starting a new session? How do you handle this today? Do you rely on Markdown files, [`AGENTS.md`](http://AGENTS.md), built-in memory, Git, custom scripts, or something else? **I’m interested in your actual experiences:** * What problems keep happening? * What existing solutions have you tried, and where do they fall short? * What do you wish worked better? I’m exploring this problem as part of product research. I’m not selling anything; I want to understand how developers actually work and what remains frustrating. I’d appreciate hearing your experiences!
document fetched 2026-10-22 00:43 · canonical https://reddit.com/r/ClaudeCode/comments/1xbq4wv/how_do_developers_using_ai_coding_agents_manage