auto compaction/flush settings?
✅Solved
hey y'all, i have a question around compaction and flushing etc
every single time i let the model auto-compact, it just... goes bananas. it seems to need to run through like, 2-3 compactions before it comes good. in the mean time, it comes back and starts working on what we were doing, with minimal context, and it just always goes... really fucking badly.
is there any well known "this compaction setting just works well" type config? currently i have this
"compaction": {
"reserveTokensFloor": 20000,
"memoryFlush": {
"enabled": true,
"softThresholdTokens": 4000,
"systemPrompt": "Session nearing compaction. Store durable memories now.",
"prompt": "Write any lasting notes to memory/YYYY-MM-DD.md; reply with NO_REPLY if nothing to store.",
},
it's just lifted straight from openclaw doco, figured it would work, it does not. i'm using opus 4.6
every single time i let the model auto-compact, it just... goes bananas. it seems to need to run through like, 2-3 compactions before it comes good. in the mean time, it comes back and starts working on what we were doing, with minimal context, and it just always goes... really fucking badly.
is there any well known "this compaction setting just works well" type config? currently i have this
"compaction": {
"reserveTokensFloor": 20000,
"memoryFlush": {
"enabled": true,
"softThresholdTokens": 4000,
"systemPrompt": "Session nearing compaction. Store durable memories now.",
"prompt": "Write any lasting notes to memory/YYYY-MM-DD.md; reply with NO_REPLY if nothing to store.",
},
it's just lifted straight from openclaw doco, figured it would work, it does not. i'm using opus 4.6
Solution
What you pasted is basically the default “pre-compaction memory flush” + the reserve floor — it doesn’t really change compaction quality, it just tries to (silently) get durable notes onto disk right before the window gets tight.
When people report “it takes 2–3 compactions before it’s sane again”, it’s usually one (or both) of these:
1) Token budgeting is impossible / too tight for the actual context window you’re on
- Auto-compaction triggers when you’re near
- OpenClaw also enforces
If your real model window is smaller than you think (or misdetected), a 20k reserve can cause early + repeated compactions.
2) The default summarizer drops “working state” (tool/file state, open TODOs, current plan)
This is exactly what
1) Turn on safeguard mode (this is the biggest practical win for “compaction went weird”):
When people report “it takes 2–3 compactions before it’s sane again”, it’s usually one (or both) of these:
1) Token budgeting is impossible / too tight for the actual context window you’re on
- Auto-compaction triggers when you’re near
contextWindow - reserveTokens- OpenClaw also enforces
reserveTokens >= reserveTokensFloorIf your real model window is smaller than you think (or misdetected), a 20k reserve can cause early + repeated compactions.
2) The default summarizer drops “working state” (tool/file state, open TODOs, current plan)
This is exactly what
agents.defaults.compaction.mode: "safeguard"“Known good” config tweaks to try
1) Turn on safeguard mode (this is the biggest practical win for “compaction went weird”):
The AI that actually does things. Emails, calendar, home automation - all from your favorite chat app. New shell, ready to help. The lobster way.
83,815Members
Resources
Was this page helpful?
