Bug type

Behavior bug (incorrect output/state without crash)

Summary

Environment

• OpenClaw: 2026.3.2
• OS: Windows 11
• Node: 24.14.0
• Ollama model: qwen3:4b (Q4_K_M, 2.5GB)
• RAM: 16GB system + 11GB swap
• GPU: RTX 2060 6GB VRAM

Summary

OpenClaw hardcodes 128K context (KvSize=262144) in the agent bootstrap and sends it directly to the Ollama runner, regardless of any config values. This causes Ollama to request 36GB of memory for a 2.5GB model, making it completely unusable on any machine with less than ~40GB RAM.

Error

model requires more system memory (34.4 GiB) than is available (17.7 GiB)

Ollama runner log confirming hardcoded value

msg=load request="{Operation:fit ... KvSize:262144 ...}"
msg="kv cache" device=CUDA0 size="36.0 GiB"
msg="model weights" device=CUDA0 size="2.3 GiB"
msg="total memory" size="39.5 GiB"

Config attempts that were all ignored

All of the following were tried and had zero effect:

"models.providers.ollama.models[].contextWindow": 4096
"models.providers.ollama.models[].numCtx": 4096
"models.providers.ollama.modelOptions.num_ctx": 4096

$env:OLLAMA_MAX_NUM_CTX = "4096"

Changing api type from openai-completions to ollama — no effect.

Workaround (working)

HTTP proxy on port 11435 that intercepts requests and rewrites num_ctx before forwarding to Ollama on 11434:

// ollama-proxy.js
const http = require('http');
const PROXY_PORT = 11435;
const OLLAMA_PORT = 11434;
const MAX_CTX = 16000; // OpenClaw minimum is 16000

const server = http.createServer((req, res) => {
  let body = '';
  req.on('data', chunk => { body += chunk.toString(); });
  req.on('end', () => {
    let modifiedBody = body;
    if (body && (req.headers['content-type'] || '').includes('application/json')) {
      try {
        const parsed = JSON.parse(body);
        if (parsed.options && parsed.options.num_ctx) {
          parsed.options.num_ctx = MAX_CTX;
        } else {
          if (!parsed.options) parsed.options = {};
          parsed.options.num_ctx = MAX_CTX;
        }
        modifiedBody = JSON.stringify(parsed);
      } catch (e) {}
    }
    const options = {
      hostname: '127.0.0.1',
      port: OLLAMA_PORT,
      path: req.url,
      method: req.method,
      headers: { ...req.headers, 'host': `127.0.0.1:${OLLAMA_PORT}`, 'content-length': Buffer.byteLength(modifiedBody) }
    };
    const proxyReq = http.request(options, (proxyRes) => {
      res.writeHead(proxyRes.statusCode, proxyRes.headers);
      proxyRes.pipe(res, { end: true });
    });
    proxyReq.on('error', (err) => { res.writeHead(502); res.end('Proxy error: ' + err.message); });
    proxyReq.write(modifiedBody);
    proxyReq.end();
  });
});

server.listen(PROXY_PORT, '127.0.0.1', () => {
  console.log(`Proxy running on ${PROXY_PORT}, forwarding to ${OLLAMA_PORT}`);
});

Then set baseUrl in openclaw.json to point to the proxy:

"models": {
  "providers": {
    "ollama": {
      "baseUrl": "http://localhost:11435"
    }
  }
}

Expected behavior

contextWindow or an equivalent config key under models.providers.ollama should control the num_ctx sent to Ollama.

Impact

This bug makes OpenClaw completely unusable with Ollama on any normal consumer PC. The workaround requires running a separate proxy process on every startup, which is not acceptable for a production setup.

Steps to reproduce

1. Install Ollama and pull any model (tested: qwen3:4b)
2. Add Ollama as a provider in openclaw.json with contextWindow: 4096
3. Start openclaw gateway
4. Send any message in TUI or Telegram
5. Observe Ollama logs

Expected behavior

OpenClaw should pass num_ctx=4096 (or whatever contextWindow is set to)
to the Ollama runner.

Actual behavior

OpenClaw sends KvSize=262144 (128K tokens) to the Ollama runner regardless
of any config. Ollama requests 36GB RAM for a 2.5GB model and fails with:
"model requires more system memory (34.4 GiB) than is available (17.7 GiB)"

OpenClaw version

2026.3.2 (build 85377a2)

Operating system

Windows 11

Install method

npm install -g openclaw

Logs, screenshots, and evidence

msg=load request="{Operation:fit KvSize:262144 ...}"
msg="kv cache" device=CUDA0 size="36.0 GiB"
msg="model weights" device=CUDA0 size="2.3 GiB"
msg="total memory" size="39.5 GiB"
msg="Load failed" error="model requires more system memory (34.4 GiB) than is available (17.7 GiB)"

Impact and severity

Blocks workflow — completely. Any user running Ollama on a machine with
less than ~40GB RAM cannot use OpenClaw with local models at all.
This affects the majority of consumer hardware (most PCs have 16-32GB RAM).
Happens every time, 100% reproducible. Workaround requires running a
separate Node.js proxy process on every startup.

Additional information

Tried all of these config keys — all ignored:

• models.providers.ollama.models[].contextWindow: 4096
• models.providers.ollama.models[].numCtx: 4096
• OLLAMA_MAX_NUM_CTX=4096 env var

Workaround: HTTP proxy on port 11435 that rewrites num_ctx in the
request body before forwarding to Ollama on 11434. See proxy code in
the body above.