#Chat

1 messages · Page 1 of 1 (latest)

pale ridge
#

Since some of the recent updates not sure if it’s related but all of sudden chat is very slow . Keeps giving me error showing like 5mil/64k context. How do I fix?

shell heronBOT
dawn tartanBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> âś… Mark Solution

shell heronBOT
obsidian matrixBOT
#

That usually means the session context blew up. 5mil/64k is way over the model window.

First, what does openclaw --version show?

Quick fixes to try now:

  1. In that chat, run /compact
  2. If it still stays slow, run /new (or /reset) for a fresh session
  3. Run /context list to see what is inflating the prompt

A useful shortcut:

  • If /new fixes it, the problem was the old session history
  • If /new does not fix it, something in the injected context is huge

If you want, paste these and I can narrow it down fast:

  • openclaw --version
  • openclaw status --all
  • /context list
pale ridge
#

that didnt work for some reason not allowing me to post screenmshots in chat

obsidian matrixBOT
#

No worries, screenshots aren’t needed.

Just paste the text, even partial. The most useful bits are:

openclaw --version
openclaw status --all

And in chat:

/context list

If Discord is being annoying, wrap the output in triple backticks so it stays readable.

#

Also try this quick test:

  • start a brand new thread/chat
  • send a normal message there

If the new thread is fast, then the old session is just bloated/corrupted and /new in that session is the fix.
If the new thread is also slow, it’s more likely a broader config/runtime issue.

If /context list will not run, just paste the exact error text you see. That’s enough for me to diagnose it.

pale ridge
#

right now its just queued thats how slow its running i let you know once it responds

obsidian matrixBOT
#

Yep, that tracks. If it’s stuck queued, I wouldn’t keep fighting that thread.

Best move now:

  1. open a fresh thread/chat
  2. send a simple test message there
  3. if that one is fast, the old session is the problem, and you can just abandon it or use /new

If the old one eventually comes back, grab the exact error text and paste it here. That should tell us whether it’s just bad session history or a real compaction bug.

pale ridge
#

🧠 Context breakdown
Workspace: /Users/pauloyler/.openclaw/workspace
Bootstrap max/file: 12,000 chars
Bootstrap max/total: 60,000 chars
Sandbox: mode=off sandboxed=false
System prompt (estimate): 45,045 chars (~11,262 tok) (Project Context 27,852 chars (~6,963 tok))

Injected workspace files:

AGENTS.md: OK | raw 4,843 chars (~1,211 tok) | injected 4,843 chars (~1,211 tok)
SOUL.md: OK | raw 5,139 chars (~1,285 tok) | injected 5,139 chars (~1,285 tok)
TOOLS.md: OK | raw 3,425 chars (~857 tok) | injected 3,425 chars (~857 tok)
IDENTITY.md: OK | raw 442 chars (~111 tok) | injected 442 chars (~111 tok)
USER.md: OK | raw 858 chars (~215 tok) | injected 858 chars (~215 tok)
BOOTSTRAP.md: OK | raw 1,450 chars (~363 tok) | injected 1,450 chars (~363 tok)
MEMORY.md: OK | raw 11,117 chars (~2,780 tok) | injected 11,117 chars (~2,780 tok)
Skills list (system prompt text): 8,710 chars (~2,178 tok) (15 skills)
Skills: canva-instagram-post, canva-poster, coding-agent, healthcheck, maintain-lead-magnet, mission-control-builder, node-connect, openai-whisper, skill-creator, taskflow, taskflow-inbox-triage, upwork-automation-using-ai, use-my-browser, video-frames, weather
Tool list (system prompt text): 0 chars (~0 tok)
Tool schemas (JSON): 13,900 chars (~3,475 tok) (counts toward context; not shown as text)
Tools: read, edit, write, exec, process, cron, sessions_list, sessions_history, sessions_send, sessions_yield, sessions_spawn, subagents, session_status, web_search, web_fetch

Session tokens (cached): unknown / ctx=64,000

Inline shortcut: a command token inside normal text (e.g. “hey /status”) that runs immediately (allowlisted senders only) and is stripped before the model sees the remaining message.

#

context breakdown

obsidian matrixBOT
#

Nice, this helps a lot.

Your system prompt looks normal. It’s only about 11k tokens, so the 5mil/64k problem is not coming from AGENTS.md, MEMORY.md, skills, etc.

That means the blow-up is almost certainly in:

  1. the session history/transcript, or
  2. a huge old tool result stuck in history

So the fix is:

  • run /new in this thread
  • send one short test message after that

If this thread is too wedged to do that, open a fresh thread/chat and use that instead.

What your numbers say:

  • MEMORY.md is a bit chunky, about 2.8k tok
#
  • skills are about 2.2k tok
  • both are fine, they add some overhead, but they do not explain a 5 million token error

If it happens again right after /new, then I’d treat it like a real bug. In that case, please paste:

openclaw --version
openclaw status --all

At this point, my bet is the old session is just cooked.