A small system that remembers what you've read, and notices when it connects.
Little Thot is a personal research assistant that runs on our own hardware. It keeps track of notes, ideas and reading, and speaks up when something new relates to something old.
It is also a laboratory. It's a real agent system, used every day, that we can take apart to understand how these systems behave.
Why we built it
Most of what we read is forgotten within a week. Notes collect in one place, bookmarks in another, half-formed ideas in a chat window. The links between them are where the interesting thinking happens, and they are the first thing to go.
There is a second reason. Agent systems are being built quickly, often faster than anyone can say how data moves through them, what they can reach, or what they cost to run. We wanted one we could see all the way through: small enough to run locally, on our own hardware, useful enough to rely on, and instrumented well enough to study.
What it does
- “You've covered this before”
- Read a page that relates to one of your notes or ideas, and it tells you, with a link back to what you wrote.
- Links you didn't ask for
- Once a week it looks across everything for connections between unrelated things, and explains why it thinks they belong together.
- “I vaguely remember…”
- Ask in plain language. It searches by meaning, so a half-remembered description is enough.
- Research on request
- Ask for background on a topic and it searches the web, frames what it finds around your own notes, and files a written report.
- An eye on the news
- It reads technology news front pages and flags the few stories that touch what you're working on.
- A profile you can correct
- It keeps a short description of what you're working on, from your own statements and a weekly reading of recent activity. You can amend it, or tell it to forget.
How it fits together
Little Thot is a handful of small programs that talk to each other through a message queue. Each does one job. Most of it runs locally, on a single-board computer. A second machine does the heavier work of turning text into vectors that can be searched by meaning.
Comes in
- Chat messages
- Pages read in the browser
- Notes, from folders it's allowed to see
- News front pages
Coordinates
- Agent reads each request, decides what it needs, and replies
- Message queue carries work between the parts
Does the work
- Capture
- Search
- Research
- Connections and the weekly digest
- Health and cost monitoring
Remembers
- One local database
- Embeddings from a local model
Choices we made
- Local first. Browsing history stays in the browser until it syncs to our own local server, and it can't sync from anywhere else. Notes are read only from folders on an allow-list.
- Small parts, one queue. The components share nothing but messages and a database, so any of them can be moved, restarted or replaced on its own.
- The big model, sparingly. Matching and search use a small embedding model on our own machine. A frontier model is called only where reasoning earns its cost.
- Everything is counted. Every model call is logged with its purpose and cost. The system checks its own health and says when something has stopped.
- A person in the loop. Every unprompted message can be marked useful or not, and those ratings are kept to tune what it says next time.
What we're studying
Little Thot is a means to an end. Running a real system every day raises questions a diagram doesn't.
- Agents
- Where should a language model make the decision, and where should ordinary code? How does a system that acts on its own avoid becoming noise?
- Data flows
- What leaves our own machines, when, and why? Can each piece of output be traced back to the inputs that produced it?
- Security
- Web pages, news stories and personal notes all end up in front of a language model. How do you contain prompt injection, keep each component's access to a minimum, and keep the damage from any one mistake small?
- Cost and reliability
- What does a personal agent actually cost to run, and what breaks first?
- Usefulness
- Which interruptions help, and which get ignored? The ratings are the measurement.
Where it is now
Little Thot is a research project, not a product, so there's nothing to sign up for. We write up what we learn in articles.