A tiny local model triggers compaction errors in every new session

Troubleshooting
, , , ,
1

Adapted community discussion: Paraphrased and condensed from an anonymized public OpenClaw Discord support conversation from August 2026. Technical guidance was reviewed for OpenClaw 2026.9.3.

My new local-model setup immediately reports that auto-compaction could not recover the turn. It recommends a larger reserve, but the same error appears in a brand-new session after only a few short messages.

2

A normal long-session compaction failure can come from the context limit, but a tiny new session points toward the local backend’s effective context or response handling. Run these on the Gateway host, not as chat messages:

openclaw --version
openclaw config get agents.defaults.compaction
openclaw models status
openclaw logs --follow

Also check the model server’s configured context length. Raising a reserve above the server’s real context cannot help.

3

Even a clean session fails. This is a small Qwen 2.5 1.5B model on a test VPS with 2 vCPU and 4 GB RAM. The hosting trial does not provide SSH, so I only have its dashboard and logs.

4

That model may fit when quantized, but it is still a weak match for OpenClaw’s system prompt, tool schemas, and compaction summaries. The failure can occur before useful conversation history accumulates.

Confirm the server’s real context setting and avoid a reserve larger than that limit. If the platform supports a lean local-model profile or tool reduction, use it for testing. Otherwise use a hosted model as the control case.

5

The hosted OpenAI model starts and answers normally. The machine has only about 1.4 GB available before loading Qwen, and the configured compaction reserve was larger than the local model’s advertised context.

6

That isolates the problem to the local backend, context setting, or available memory rather than the Gateway as a whole. Lower the reserve to fit within the model’s real context, keep enough RAM for the KV cache and OpenClaw runtime, and test the local model independently. If the platform cannot expose server logs or context controls, its trial environment is not sufficient for diagnosing this setup.