Cline Desktop v0.0.19 Fixes Memory Leak That Ballooned Process to Tens of Gigabytes

What shipped

Cline released desktop v0.0.19 on August 26, fixing a memory leak in the background Cline process and refreshing its model provider catalog.

The memory bug

According to the release notes, session status updates were carrying a full copy of the conversation transcript to every connected client. On multi-megabyte tasks, this caused the background process to grow to tens of gigabytes of memory during long sessions.

The fix changes what status updates carry. They now include only state data — status, usage, model, workspace, and checkpoint — rather than the full transcript. The transcript itself is now fetched on demand instead of being pushed with every update. This is a structural fix to the client-server communication pattern, not a tuning adjustment, so it should eliminate the scaling problem rather than just delay it.

Provider catalog changes

The release adds seven new model providers: Agnes AI, Aixy, IteraCompute, LLM Tech, NeoSmith, Pendra, and Standard Compute. None of these appear in major model registries as of this release, suggesting they are smaller or newer inference providers being aggregated into Cline's routing layer.

Cline also updated model lists and pricing across existing providers. Notably, the resolved default model changes for several providers if users have not pinned a specific model:

• ClinePass: now defaults to GLM 5.3
• Z.ai
• Hugging Face
• evroc
• LLM Gateway
• NanoGPT
• Weights & Biases

Users on these providers who rely on Cline's default model selection will see a different model served after upgrading, even though they made no configuration change themselves.

What this means

The memory leak fix addresses a real reliability problem for anyone running long-lived Cline sessions — a process ballooning to tens of gigabytes would eventually crash or degrade the host machine, which matters for developers running agentic coding sessions unattended for hours. The fix itself (sending state instead of full transcripts) is a sensible architectural correction rather than a workaround.

The provider catalog expansion signals that Cline is functioning increasingly as a model-agnostic routing layer, aggregating an expanding and fragmented set of inference providers rather than betting on a small number of large labs. For users, the practical takeaway is to check pinned model configurations before upgrading: if you were relying on a default rather than an explicit model selection on ClinePass, Z.ai, Hugging Face, evroc, LLM Gateway, NanoGPT, or Weights & Biases, your requests will now route to a different model with potentially different cost and quality characteristics.