Help needed: OpenClaw agent maxing out 200k context window | Friends of the Crustacean 🦞🤝 | Page 1

warm yarrowBOT Apr 24, 2026, 9:48 AM

#

finite stirrupBOT Apr 24, 2026, 9:48 AM

#

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

warm yarrowBOT Apr 24, 2026, 9:48 AM

#

fair fjordBOT Apr 24, 2026, 9:50 AM

#

I’m checking the docs/source for how OpenClaw compacts history and whether this looks like a model/provider issue vs a config issue.

#

That sounds annoying. I think this is most likely tool-output bloat, not “200k is too small”.

A few concrete things to check first:

Paste your version
```
openclaw --version
```
Paste a status snapshot
```
openclaw status --all
```
In the chat/session that keeps bloating, run:

#

   /status
   /context detail

/context detail is the big one, it shows what is actually consuming tokens.

Paste these config bits from openclaw.json:
- models.providers.ollama.baseUrl
- agents.defaults.contextPruning
- agents.defaults.compaction

Two likely causes from the docs:

Session pruning is off by default for non-Anthropic providers, so with Ollama/Kimi, old tool results can pile up fast.
If your Ollama base URL uses /v1, that is wrong for OpenClaw and can break tool calling / dump raw tool JSON into history. It should be the native Ollama URL, for example:

#

  {
    models: {
      providers: {
        ollama: {
          baseUrl: "http://your-ollama-host:11434"
        }
      }
    }
  }

If pruning is missing, I’d try this first:

#

  agents: {
    defaults: {
      contextPruning: {
        mode: "cache-ttl",
        ttl: "5m",
        keepLastAssistants: 3,
        softTrimRatio: 0.3,
        hardClearRatio: 0.5,
        minPrunableToolChars: 50000,
        softTrim: { maxChars: 4000, headChars: 1500, tailChars: 1500 },
        hardClear: { enabled: true, placeholder: "[Old tool result content cleared]" }
      }
    }
  }
}

#

Also useful right now:

/compact

If the session is already cursed, just start fresh:

/new

Relevant docs:

#

If you paste the version, status, /context detail, and those config bits, I can usually pinpoint it pretty fast.

half atlas Apr 24, 2026, 10:49 AM

#

OpenClaw 2026.4.22 (00bd2cf)
Overview
┌─────────────────┬────────────────────────────────────────────────────────────────────────────────────────────────────┐
│ Item │ Value │
├─────────────────┼────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ Version │ 2026.4.22 │
│ OS │ linux 6.8.0-110-generic (x64) │
│ Node │ 22.22.2 │
│ Config │ ~/.openclaw/openclaw.json │
│ Dashboard │ http://127.0.0.1:18789/ │
│ Tailscale │ off │
│ Channel │ stable (config) │
│ Update │ pnpm · up to date · npm latest 2026.4.22 │
│ Gateway │ local · ws://127.0.0.1:18789 (local loopback) · reachable 75ms · auth token · af853a79e598 (172. │
│ │ 16.1.2) app 2026.4.22 linux 6.8.0-110-generic │
│ Security │ Run: openclaw security audit --deep │
│ Gateway self │ unknown │
│ Gateway service │ systemd not installed │

#

Node service │ systemd not installed │
│ Agents │ 1 total · 0 bootstrapping · 0 active · 2 sessions │
│ Secrets │ none

#

Status:
🦞 OpenClaw 2026.4.22 (00bd2cf)
🧠 Model: ollama/kimi-k2.6:cloud · 🔑 unknown
🧮 Tokens: 235k in / 670 out
📚 Context: 59k/200k (30%) · 🧹 Compactions: 3
🧵 Session: agent:main:main • updated just now
⚙️ Runtime: direct · Runner: pi (embedded) · Think: low
🪢 Queue: collect (depth 0)

Context detail:
🧠 Context breakdown (detailed)
Workspace: /data/.openclaw/workspace
Bootstrap max/file: 12,000 chars
Bootstrap max/total: 60,000 chars
Sandbox: mode=off sandboxed=false
System prompt (run): 35,729 chars (~8,933 tok) (Project Context 16,408 chars (~4,102 tok))
Injected workspace files:
• AGENTS.md: OK | raw 7,907 chars (~1,977 tok) | injected 7,907 chars (~1,977 tok)
• SOUL.md: OK | raw 2,677 chars (~670 tok) | injected 2,677 chars (~670 tok)
• TOOLS.md: OK | raw 952 chars (~238 tok) | injected 952 chars (~238 tok)
• IDENTITY.md: OK | raw 504 chars (~126 tok) | injected 504 chars (~126 tok)
• USER.md: OK | raw 905 chars (~227 tok) | injected 905 chars (~227 tok)
• HEARTBEAT.md: OK | raw 1,438 chars (~360 tok) | injected 1,438 chars (~360 tok)

#

• BOOTSTRAP.md: MISSING | raw 0 | injected 0
• MEMORY.md: OK | raw 3,203 chars (~801 tok) | injected 3,203 chars (~801 tok)
Skills list (system prompt text): 4,349 chars (~1,088 tok) (8 skills)
Skills: clawhub, healthcheck, himalaya, node-connect, skill-creator, taskflow, taskflow-inbox-triage, weather
Top skills (prompt entry size):
• skill-creator: 756 chars (~189 tok)
• taskflow: 540 chars (~135 tok)
• node-connect: 538 chars (~135 tok)
• healthcheck: 488 chars (~122 tok)
• clawhub: 429 chars (~108 tok)
• taskflow-inbox-triage: 415 chars (~104 tok)
• weather: 413 chars (~104 tok)
• himalaya: 380 chars (~95 tok)
Tool list (system prompt text): 0 chars (~0 tok)
Tool schemas (JSON): 25,136 chars (~6,284 tok) (counts toward context; not shown as text)
Tools: read, edit, write, exec, process, canvas, nodes, cron, message, tts, gateway, agents_list, sessions_list, sessions_history, sessions_send, sessions_spawn, sessions_yield, subagents, session_status, web_fetch, browser, memory_search, memory_get
Top tools (schema size):
• cron: 6,700 chars (~1,675 tok)
• message: 5,767 chars (~1,442 tok)
• browser: 2,799 chars (~700 tok)
• nodes: 1,589 chars (~398 tok)
• sessions_spawn: 1,333 chars (~334 tok)
• exec: 1,096 chars (~274 tok)
• process: 961 chars (~241 tok)
• edit: 834 chars (~209 tok)
• canvas: 661 chars (~166 tok)
• gateway: 537 chars (~135 tok)
• sessions_list: 432 chars (~108 tok)
• web_fetch: 374 chars (~94 tok)
• read: 304 chars (~76 tok)

fair fjordBOT Apr 24, 2026, 10:51 AM

#

Nice, version looks current, so this is probably not a stale-build bug.

What I’d check next:

Make sure Ollama is using the native API URL, not /v1
```
{
  models: {
    providers: {
      ollama: {
        baseUrl: "http://your-ollama-host:11434"
      }
    }
  }
}
```
If you use ...:11434/v1, OpenClaw docs call that out as bad for tool calling, and it can dump ugly raw tool JSON into history.

#

Docs: https://docs.openclaw.ai/providers/ollama

Enable session pruning explicitly
For non-Anthropic setups, this is the first thing I’d try:

{
  agents: {
    defaults: {
      contextPruning: {
        mode: "cache-ttl",
        ttl: "5m",
        keepLastAssistants: 3,
        softTrimRatio: 0.3,
        hardClearRatio: 0.5,
        minPrunableToolChars: 50000,
        softTrim: { maxChars: 4000, headChars: 1500, tailChars: 1500 },

#

           hardClear: { enabled: true, placeholder: "[Old tool result content cleared]" }
         }
       }
     }
   }

This trims old tool results before each LLM call, which is usually the real context killer.
Docs: https://docs.openclaw.ai/concepts/session-pruning

Then start a fresh session
```
/new
```
If the current thread is already bloated, fixing config alone won’t magically make that history small.
Paste this too

#

   /context detail

That’s the fastest way to see whether the growth is:

tool results
giant bootstrap/system prompt
attachments/images
just a runaway long transcript

If you want, paste these 3 config sections and I can sanity-check them:

models.providers.ollama
agents.defaults.contextPruning
agents.defaults.compaction

Also useful docs:

https://docs.openclaw.ai/concepts/compaction

#

https://docs.openclaw.ai/reference/token-use

My bet right now: either /v1 is in the Ollama URL, or pruning is off, and old tool output is eating the whole window.

half atlas Apr 24, 2026, 10:53 AM

#

"providers": {
"ollama": {
"baseUrl": "http://172.16.1.1:32769",
"api": "ollama",
"apiKey": "ollama-local",
"models": [
{
"id": "kimi-k2.6:cloud",
"name": "kimi-k2.6:cloud",
"api": "ollama",
"reasoning": false,
"input": [
"text"
],
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0
},
"contextWindow": 200000,
"maxTokens": 8192
},
{
"id": "nomic-embed-text",
"name": "nomic-embed-text",
"api": "ollama",
"reasoning": false,
"input": [
"text"
],
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0
},
"contextWindow": 2048,
"maxTokens": 2048

"defaults": {
"workspace": "/data/.openclaw/workspace",
"model": {
"primary": "ollama/kimi-k2.6:cloud"
},
"models": {
"ollama/kimi-k2.6:cloud": {
"alias": "kimi-k2.6"
}
},
"thinkingDefault": "low",
"memorySearch": {
"enabled": true,
"sources": [
"memory",
"sessions"
],
"experimental": {
"sessionMemory": true
},
"provider": "ollama"
},
"compaction": {
"memoryFlush": {
"enabled": true
}
},
"maxConcurrent": 4,
"subagents": {
"maxConcurrent": 8
}
}
},

fair fjordBOT Apr 24, 2026, 10:54 AM

#

Yep, this helps a lot.

Good news: your Ollama config looks correct. You are using the native API (api: "ollama", no /v1), so I don’t think this is the /v1 tool-calling bug.

What jumps out instead: you have no agents.defaults.contextPruning configured at all.

For this setup, that’s the first thing I’d fix. Without pruning, old tool results just keep piling up in the in-memory prompt until compaction has to save you.

I’d change this first:

{
  agents: {
    defaults: {
      contextPruning: {
        mode: "cache-ttl",

#

        ttl: "5m",
        keepLastAssistants: 3,
        softTrimRatio: 0.3,
        hardClearRatio: 0.5,
        minPrunableToolChars: 50000,
        softTrim: {
          maxChars: 4000,
          headChars: 1500,
          tailChars: 1500
        },
        hardClear: {
          enabled: true,
          placeholder: "[Old tool result content cleared]"
        }
      }

#

    }
  }
}

Second thing, a bit suspicious:

memorySearch: {
  enabled: true,
  sources: ["memory", "sessions"],
  experimental: { sessionMemory: true }
}

That is not automatically bad, but if the agent likes calling memory_search, it can start pulling transcript-derived stuff back into the conversation. With pruning off, that can get ugly fast.

#

So while debugging, I’d temporarily simplify it to:

memorySearch: {
  enabled: true,
  sources: ["memory"],
  experimental: { sessionMemory: false },
  provider: "ollama"
}

Then:

restart OpenClaw
start a fresh chat session with:

#

   /new

run:
```
/context detail
```

If it still blows up after that, paste the /context detail output. That report is the fastest way to see whether the bloat is coming from:

memory_search
read
exec
huge tool schemas/results
or just plain transcript length

Docs:

https://docs.openclaw.ai/concepts/session-pruning

#Help needed: OpenClaw agent maxing out 200k context window