February 20, 2026
Memory Yeet: AI ate your DOM
Claude Code's compaction discards data that's still on disk
Users say Claude “forgets” your work mid-task, cue chaos
TLDR: Claude Code’s auto-compaction can drop data you’re using, even though it remains saved on your computer. The community is split between scrapping compaction and fixing it with smart pointers; everyone agrees lost context fuels hallucinations and wasted time, which is a big deal for reliable coding assistants.
Claude Code just got dragged for a messier kind of “clean up.” Auto-compaction—meant to shrink long chats—allegedly tosses out the very data users are working on, leaving only a vague note like “user provided DOM markup.” One dev says compaction fired mid-session and suddenly the bot started guessing at details from 20 minutes earlier while the real text sat untouched on disk. The mood? AI amnesia jokes and “did my bot just ghost my code?” vibes.
The thread split fast. One camp, led by voices like CjHuber, says ditch compaction entirely—why risk lossy summaries when hallucinations are already a problem? Another chorus, channeling selridge’s “chat is exhaust,” says: stop trusting the chat bubble; save everything you care about. Meanwhile, a side quest breaks out as someone asks how to make Claude talk through a proxy—because of course. On the flip side, SatvikBeri calls compaction critical and shouts out rival tool Amp’s smoother “handoff,” intensifying the tool-war energy.
The proposed fix is getting cheers: tag compacted notes with line numbers so Claude can fetch the original text on demand—like a surgical “undo” that costs tokens only when needed. With multiple open issues and hours of work at stake, the community wants receipts, not vibes.
Key Points
- •Auto-compaction in Claude Code removes user-provided content from working context while leaving the full transcript on disk.
- •Compacted summaries retain only generic references (e.g., “user provided DOM markup”) without pointers to the original data.
- •This issue affects tasks with large inputs and is reflected in multiple open issues reporting related symptoms.
- •Root cause: compaction is a one-way, lossy process that is not connected back to the lossless on-disk transcript.
- •Proposed fix: add transcript line-range annotations in summaries to enable on-demand, precise recovery, with a future cross-session lookup extension.