Local Ollama subagents failing with 8K context

[dusky hinge]

I am new to using Openclaw -->my local Ollama subagents failing with 8K context limit even after setting OLLAMA_NUM_CTX=32768 and num_ctx 32768 in Modelfile — anyone solved this?"

[deep perch]

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

[dusky hinge]

Local Ollama subagents failing with 8K context

[worn wagon]

First thing: please paste openclaw --version and ollama --version — version mismatches matter a lot here.

Likely fix: for stock Ollama, OLLAMA_NUM_CTX isn’t the documented env var; use OLLAMA_CONTEXT_LENGTH=32768 when starting ollama serve, or send num_ctx via the API/config. (github.com)

Also make sure OpenClaw’s config budget and Ollama’s runtime budget both say 32768:

models: {
  providers: {
    ollama: {
      baseUrl: "http://127.0.0.1:11434", // no /v1
      apiKey: "ollama-local",
      api: "ollama",
      contextWindow: 32768,
      models: [{
        id: "your-model",

        name: "your-model",
        contextWindow: 32768,
        maxTokens: 8192,
        params: { num_ctx: 32768, keep_alive: "15m" },
      }],
    },
  },
}

OpenClaw docs specifically say contextWindow is OpenClaw’s prompt/compaction budget, while params.num_ctx is what gets sent to Ollama; keep them aligned. Also don’t use http://host:11434/v1 for Ollama in OpenClaw unless you intentionally want the less reliable OpenAI-compatible path. (docs.openclaw.ai)

If it only fails for subagents, check they’re actually using this model: subagents inherit the caller unless agents.defaults.subagents.model / per-agent subagent model is set. (docs.openclaw.ai)

Useful pasteables:

openclaw models list --provider ollama
openclaw models status
openclaw config get models.providers.ollama
openclaw config get agents.defaults.subagents.model

For “has anyone solved this on similar hardware?” also ask in #users-helping-users — that’s better for real-world Ollama/GPU experiences.

[dusky hinge]

sorry for the late response. my OpenClaw version is 2026.4.22 and i am running Ollama at 0.18.3.

[worn wagon]

No worries — thanks, that helps.

You’re a few releases behind. I’d update OpenClaw first; there have been Ollama/subagent fixes since 2026.4.22.

For your exact version, the likely issue is: OpenClaw 2026.4.22 may still be seeing the model as 8K and sending that to Ollama, regardless of your OLLAMA_NUM_CTX / Modelfile attempt.

Do this:

openclaw update
openclaw doctor
openclaw gateway restart

Then verify what OpenClaw thinks the model context is:

openclaw models list --provider ollama

If it still shows 8K, define the Ollama model explicitly in ~/.openclaw/openclaw.json:

{
  models: {
    providers: {
      ollama: {
        baseUrl: "http://127.0.0.1:11434",
        apiKey: "ollama-local",
        api: "ollama",
        contextWindow: 32768,
        maxTokens: 8192,

        models: [
          {
            id: "your-model-name",
            name: "your-model-name",
            input: ["text"],
            contextWindow: 32768,
            params: { num_ctx: 32768 }
          }
        ]
      }
    }
  },
  agents: {
    defaults: {
      model: { primary: "ollama/your-model-name" }

    }
  }
}

Important bits:

• use baseUrl: "http://127.0.0.1:11434" — no /v1
• use api: "ollama" for native Ollama mode
• OLLAMA_NUM_CTX is not the one I’d rely on; Ollama’s documented env is OLLAMA_CONTEXT_LENGTH
• after config changes: openclaw gateway restart and start a fresh session/thread if possible

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

If it still fails, paste the exact error plus:

openclaw status --all
openclaw models list --provider ollama

[dusky hinge]

Thanks.. this helps