mindfork mindfork

mindfork 0.17.0 — your Ollama or LM Studio, found and offered

· 3 min read

mindfork 0.17.0 is out. If you already run Ollama or LM Studio, connecting mindfork to it is now one key.

Found and offered

Start mindfork with no model connected, and it looks at this computer for an Ollama and an LM Studio. What answers is listed: one row per server and model, a loaded model first.

  • Enter connects the row. It is written into the settings the way the settings screen writes it, so nothing is hidden: Model/server shows it, and changes it.
  • The embedding model comes with it, when none is set up yet. LM Studio ships one, and Ollama serves any you have pulled. Notes and the knowledge base then search by meaning from the first message. An embedding model you already set up is never replaced: a different one would mean indexing everything again.
  • Esc changes nothing. The list comes back at the next start for as long as nothing is connected.
  • /local looks again whenever you type it — after starting a server later, or to switch to one. When nothing answers, it says which ports it asked.

LM Studio's server is off until you start it: lms server start, or the Developer tab of the app. An Ollama with no model needs one pulled first.

LM Studio, a server mindfork knows

LM Studio already worked with mindfork as any OpenAI-compatible server does. Run live, it passed everything but one check: mindfork did not know its window. Now it reads it from LM Studio, so a long conversation is folded into a summary before LM Studio would cut it.

And LM Studio does cut. Measured on LM Studio 1.1.7, it answers a prompt longer than its window in one of two ways:

  • When your last message alone does not fit beside mindfork's instructions, it refuses. mindfork says the conversation no longer fits.
  • Otherwise it drops the middle of the conversation without saying so: it keeps the instructions, the first message and the last, and everything in between goes at once. In the test, a conversation of 4044 tokens was answered with every turn remembered. One turn more, and LM Studio processed 2809 tokens and the model remembered only the first.

mindfork tells such a cut from the size LM Studio reports, as it already did for Ollama since 0.16.1.

The notes speak of the server

When the window is too small, or a prompt was cut, the note now names the setting of the server you run:

  • Ollama: OLLAMA_CONTEXT_LENGTH, or the context slider in its app.
  • LM Studio: the model's Context Length, or lms unload <model> and then lms load <model> -c 16384. The unload matters: loading a loaded model again starts a second copy, and mindfork's requests keep going to the first.
  • Any other server: its own settings, with no product named. Until now, every server was told about Ollama's.

Your data

Nothing in the data's format changes, and no step runs at the start. A server you pick is written into the settings you could have typed by hand.

The full list is in the changelog; downloads are on the releases page, and cargo install mindfork builds the same version from crates.io.