Work that compounds

Stop restarting your research from scratch every session

A chat thread is a conversation you have once. Agent Bayes is a workspace you return to, where each session builds on the last one: the same mindmap, the same conversation history, version history behind it, and snapshots of every state worth keeping.

Create your account

Free to start, no card. 200 credits when you verify your email.

The reason you are here

You have explained your project to a chatbot four times

Each new session starts with the same paragraph of background. Which literature, which framing, which terms mean something specific in your subfield. By the time the model is useful, you have spent twenty minutes rebuilding a context that existed last Tuesday.

Then the useful part arrives as prose in a scrolling thread, and a week later you cannot find the exchange where it happened, or tell which version of your thinking it belonged to.

Research accumulates. A tool that resets every session is working against the shape of the work.

How it accumulates

One map, six weeks, nothing re-explained

The artifact grows. The context travels with it.

Week one: the first pass
Index a handful of papers and ask your opening question. You get a small map of cited claims, which is more than a chat gives you and less than you will eventually need.
Week three: it has grown
New papers, new branches, a few nodes you rewrote yourself. The conversation history explains why each branch exists, including the reasoning you would otherwise have forgotten.
Week six: you write from it
The map is now the thing you draft against. Snapshots mark the states worth returning to, and every claim still points at a page in a paper.

In the product

The history is a first-class part of the workspace

Conversations are scoped to a mindmap, and each node carries the record of how it got there.

The conversations list for a mindmap, showing past agent conversations with their history — Every conversation stays attached to the mindmap it shaped.

The node properties panel showing model, sources, confidence and edit history for a claim — Per-node provenance: what wrote it, what it drew on, and what changed since.

What you can rely on

Six things that persist

Conversation history per mindmap

Every mindmap keeps its own agent conversations in full. Come back in three weeks and the thread is still there, attached to the map it produced.

Twenty recent versions, always

A rolling stack of recent versions gives you real undo and redo across sessions, not just within one.

Named snapshots that never expire

Save a known-good state before a big restructure. Snapshots live outside the rolling stack, can be diffed against the current map, and restore as a new version, so nothing is overwritten. Saving and restoring is free.

Replay what the agent changed

Additions, deletions and word-level rewrites highlighted on the canvas. A review toggle shows what an undo would revert before you commit to it.

Many maps per project

One project holds a mindmap per research question, each with its own history, so a side question does not contaminate the main line of work.

Runs survive you leaving

Navigate away mid-run and the agent keeps working on durable infrastructure. Reconnect later to a live stream of its progress, or cancel and pay only for what ran.

What Agent Bayes does not do

It is not a note-taking app. If you want free-form notes and daily journals, keep the tool you already use for that.
It does not replace your reference manager. Zotero stays the source of truth for your library, and the plugin works from it.
It does not run unattended. The map grows because you direct it, session after session, and it stops growing when you stop.

The last one is the honest version of the pitch. This compounds because you keep returning to it, not because it works while you sleep.

Before you ask

Questions researchers ask first

What happens if the agent makes a mess of my map?

Undo it. A rolling stack of the last twenty versions covers ordinary mistakes, and a named snapshot covers the big ones. Restoring creates a new version rather than overwriting, so the state you restored from is still there.

Can I see what changed after a run?

Yes. Replay the edit visually on the canvas with additions, deletions and word-level rewrites highlighted, and toggle a review view that shows what an undo would revert.

Do I lose the run if I close the tab?

No. Workflows run on durable infrastructure, so the agent keeps going. Reconnect and you pick up a live stream of its progress. Cancel any time and you are charged only for what actually ran.

Is provenance kept for my own edits too?

Yes. Each node records which model wrote it or whether you did, which sources it drew on, its confidence score, and timestamps for every change.

What does it cost to try?

The free plan costs nothing and gives you 200 one-time credits once you verify your email, enough for about five indexed papers and a short piece of writing. Saving and restoring snapshots costs nothing.