Pi + Ollama - context lost with every prompt

[darrendavid]

Kicking the tires on Pi and I'm looking to set up a self-hosted agentic coding workflow. So far, Pi works great with Claude Code, and I also have Ollama running on a separate machine on my LAN. I used @kylebrodeur/pi-model-discovery to sync my models and select devstral-small-2 running on an ADA 4500 (24GB). In my Pi footer I can see the model selection and my 393k default context window is accurately reflected. However, when start a thread with a prompt, my question gets correctly answered, but for all subsequent replies all context is lost and the model basically tells me "I don't know what you're referring to." Same issue with other models. These models (and all others) work fine in the same Ollama host via an Open Web UI chat, so I'm wondering if I've missed a critical step in Pi setup.

I've been relying on the Claude Code CLI for so long and this is my first "DIY" harness, so perhaps I've omitted a critical extension or package to achieve parity in my experience. Any obvious things come to mind? Happy to RTFM if a link will set me straight.

Thank you!

[darrendavid]

Well, further investigation seems to reveal that it's also losing context in Open Web UI, so there's clearly something else off here. Will dig in an report back when I know more in the hopes I can help someone else.

[Pxtl]

Ollama often has incredibly small context windows by default on models. When I used qwen I immediately had to create a custom model wrapping it within Ollama to give it enough context-window to be useful.

For example (link to my repo that I use to store my local agent+harness):

https://github.com/Pxtl/dirtclod-pi/blob/main/initcontent/ollama-defaults/modelfiles/qwen3.5-dirtclod-agent/Modelfile