Automatic compaction at low context? Time based?
retained text
I've experienced three auto-compactions today and much of my project engineering has been all about being strategic with my sessions, what gets loaded, layered project docs and things that load when needed, etc. I typically try to keep my sessions below 500k, and ideally below 350k. However, today's compactions seemed to be occurring based on a time window. My Claude agent summarized what happened, as in the table below: |Session|Compacted at|Size then|Your last reply before it| |:-|:-|:-|:-| |This one|10:41 pm|314k tokens (31% of 1M)|9:36 pm, **65 min earlier**| |Deltas review|10:35 pm|210k tokens (21% of 1M)|9:20 pm, **75 min earlier**| Basically, I went to walk the dog while my sessions were working on jobs, came back to a bunch of decisions to be made, communicated them in a 1-page answer dump back to Claude, and immediately got hit with an auto-compaction message and "we saved you 200k tokens" or something similar. What's going on? My session recommended that I set "autoCompactEnabled": false in \~/.claude.json Anyone else experience this? Am I doing something wrong?
document fetched 2026-10-30 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xiac7u/automatic_compaction_at_low_context_time_based