🦙💬 Ollama Talk

This is the flagship node of the "AI" side of the Bjornulf pack - a full local chat node wired into Ollama, with actual conversational memory, not just a one-shot prompt-in-text-out call. Ask it something, then ask a follow-up that only makes sense in context ("where is that?"), and it correctly resolves "that" to what you talked about two messages ago. It's the node the author leads with in his own demo videos, and community feedback on it has been genuinely warm - people specifically call out the Ollama and loop nodes together as the most powerful part of the pack.

How it works

By default it talks to llama3.2:3b on http://0.0.0.0:11434 - override either with the pack's Ollama Configuration node if you're running a bigger model or a remote Ollama server. Context is the interesting part: wire the updated_context output back into the context input of the next call in the chain and the conversation accumulates naturally. If you'd rather not manage that wiring by hand, turn on use_context_file and the node persists the running conversation to ComfyUI/Bjornulf/ollama_context.txt on disk instead - re-run the same workflow repeatedly and the conversation just keeps going, no extra nodes required. Hitting the node's reset button archives that file to ollama_context_001.txt, _002.txt, and so on, so old conversations aren't lost, just rotated out.

The other genuinely useful mode is waiting_for_prompt: turn it on and the node pauses the workflow mid-run with a Resume button, letting you edit the prompt live before it actually sends the request - an interactive back-and-forth instead of a fire-and-forget call.

One practical warning if you rely on the context-file mode: it writes to a path relative to your ComfyUI installation folder (ComfyUI/Bjornulf/...). If you installed ComfyUI through a one-click launcher or bundler that restructures the folder layout differently from a plain git checkout, that relative path can land somewhere unexpected or fail outright - worth checking that file actually exists where you expect after your first run if context isn't persisting the way you think it should.

The inputs and outputs that matter

• user_prompt - what you're asking. The seed widget next to it gets ComfyUI's usual fixed/increment/decrement/randomize control, so you can force a fresh call even when nothing else about the node changed.
• max_tokens (default 600, up to 4096) - caps response length; a token is roughly three English characters, so 600 is a fairly short answer.
• vram_retention_minutes (0–99) - keeps the model loaded in VRAM between calls for faster chaining, at the cost of holding that VRAM the whole time. Don't combine a long retention window with an image generation running in the same workflow.
• answer_single_line - forces the reply onto one line, useful if you're feeding it into a field that breaks on newlines.
• waiting_for_prompt - pauses the node for a live-edit-then-resume interactive exchange.
• use_context_file - persists conversation history to disk instead of requiring manual context wiring.
• Optional: OLLAMA_CONFIG (model/URL override), context (manual conversation history), OLLAMA_JOB (a preset task/persona from the pack's Job Selector node).
• Outputs: ollama_response (the reply), updated_context (feed this back in for the next turn), system_prompt (what system prompt actually got used).

How to install it

Via ComfyUI Manager, search Bjornulf_custom_nodes and install, or clone manually:

cd ComfyUI/custom_nodes
git clone https://github.com/justUmen/Bjornulf_custom_nodes

pip install -r requirements.txt from inside the folder (venv recommended), and separately - outside ComfyUI entirely - install and run Ollama itself from ollama.com, then ollama pull llama3.2:3b (or your model of choice). The node's own Python dependency is just pip install ollama, but that's required even if you never touch this node, since the pack imports it at load time regardless.

Common issues

Connection refused is the most common error, and it's almost always the default http://0.0.0.0:11434 not resolving to wherever Ollama is actually running - point an Ollama Configuration node at the right address if ComfyUI and Ollama aren't on the exact same machine. If you're getting a suspiciously identical response run after run and nothing about your prompt changed, set the seed widget's control to "randomize" rather than "fixed" - that's the standard fix across this whole pack for forcing a node to actually re-execute instead of ComfyUI reusing a cached result.