Colloquy

Build a knowledge base out of your AI conversations.

Each one is saved to your vault as markdown, then summarised so the next starts with the context. Supports Ollama, OpenAI, Anthropic, Google and DeepSeek.

The conversations are written as they happen: searchable, linkable, and still readable if this plugin goes away. A few minutes after each one, the plugin reads it back and keeps a short account of what you were doing, and that account is what later conversations are given as background. It reads what it wrote; it does not read the rest of your vault.

Two things it does that you should know before installing

It writes to your vault on its own. A few minutes after a conversation goes quiet, the plugin reads it and files a short account of what you were doing, not what the model claimed, what you were doing, in a folder you choose. That is the feature, and it is the most surprising thing here, so it is at the top rather than under Advanced. It can be turned off.

It sends earlier conversations with your question. Whatever it noticed about your last few conversations rides in front of the next one, so the model has context. It goes to whichever provider you have selected, including a paid one. This can be turned off too, and the setting says plainly where the text goes.

The automatic reading runs on a model that costs nothing. Reading every conversation through a metered API is a bill that grows while you sleep, so it was designed out rather than warned about. That work has its own setting: point it at a small local model and the record keeps itself while you talk to whatever you like, paid or not.

What it looks like in your vault

Conversations/2026/08/19/is-it-better-by-train.md      the conversation, as markdown
Conversations/context/2026/08/19/is-it-better…md       what it noticed about you

Both folders are settings. The defaults are plain names; if your vault has its own scheme, put your own paths in and the plugin will use them. Nothing else is ever written anywhere.

Getting started

With a local model. Nothing leaves your machine and nothing costs anything.

1. Install Ollama and start it.
2. Settings → Colloquy → download a model. gemma3:4b is a good first one: it reads an image, and it summarises a conversation in about two seconds.
3. Open the panel from the ribbon and ask something.

With an API key, better answers, and you pay per question.

1. Settings → Colloquy → paste a key for OpenAI, Anthropic, Google or DeepSeek.
2. Pick the provider and model in the chips under the composer.
3. Worth doing once: Settings → Colloquy → Model for reading and naming → Ollama. The answers come from what you are paying for, the record keeps itself for nothing.

Keys are stored in the plugin's own data file, which is never rendered back to the screen and never leaves your machine except as the request you asked for.

Using it

• Ask: type, press Enter. The answer streams in and the file is written as it goes.
• Stop: Send becomes Stop while an answer arrives. What arrived is kept.
• Link: put a URL in your question and the page is read and sent with it. A video gives its title and description, not what it contains: that page is a script bundle, and no amount of stripping tags gets a transcript out of it.
• Attach: paste a screenshot straight into the box, or use the paperclip for a note or an image, from your vault or from your computer. Anything from outside the vault is copied in first, so the conversation links to something that is actually there.
• Pick up where you left off: the history button lists past conversations. They are rebuilt from the file, so one you edited by hand resumes just the same.

What it will not do

• PDFs are refused. They need extracting rather than encoding, and sending bytes a model will silently ignore is worse than saying no.
• Resuming does not re-send attachments. The links stay in the transcript; the model picking it up sees the words, not the picture.
• A small local model gets things wrong. The accounts it writes are one model's reading of a conversation. They live in a log, are marked author: agent, and are handed back as a hint that says it may be wrong, never as fact.
• A phone needs an API key. Everything works there, the panel, attachments, resuming, the conversations in your vault, but a phone has no localhost for Ollama to be on, so the free local route is a desktop one. Point it at Ollama on your own machine over the network and that works too (ADR-0012).
• Nothing is read automatically on a phone. The unprompted reading runs on a provider that charges nothing per request, which on a phone means one reached over your network. Conversations you have on a phone are read the next time the desktop is open, unless you have deliberately named a paid model for that work.

Why it works the way it does

decisions/ holds the reasoning, one file per decision, including the ones that were reversed. ADR-0007 explains why an earlier version that extracted "durable notes" was removed entirely: it kept producing confident, false claims.

What changed when

CHANGELOG.md, one entry per release.

Building it

npm install
npm test
node build.mjs --vault "/path/to/a/test/vault"

No runtime dependencies. One build dependency, esbuild, and the build script is short enough to read in full.

Support

Colloquy is free and stays free. If it has been useful, a coffee is a kind way to say so.

Licence

MIT.