Alyph · Branching
Branch any conversation. Keep every context clean.
In a linear chat, every tangent piles onto the same history. In Alyph, any message can start a new branch that inherits only the context above it — so experiments, rabbit holes, and bake-offs never touch your main thread.
The short answer
Context branching is a conversation structure where each new thread inherits only the context of the point it branches from — the conversational equivalent of git branches. Shared context (files, constraints, decisions) lives at the root; everything below the branch point is private to that branch.
Should I take the Acme retainer, or keep the slot open?
Take it, but cap the hours. Retainers creep.
Take it if you can name the client you’d drop for it. ▍
What you get
- Start a branch from any message, at any point
- Branches inherit the root — files, constraints, decisions — and nothing else
- Sibling branches stay completely isolated from each other
- You pay for the root plus one branch, not the whole history
- Works with every model in the wallet, side by side
- Done with a branch? Prune it in one click
The mental model: git for conversations
You would never build two features by committing both to the same branch in alternating commits. A linear AI chat asks exactly that: one shared history where every task, tangent, and failed attempt piles onto the same log.
Branching applies the git model to conversations. The root holds your shared foundation — files, constraints, the original question. A branch inherits the root plus its own messages. You do the merging yourself, by taking the winning answer back into your real work. A branch that goes nowhere gets pruned.
What the model actually receives
When you send a message inside a branch, the model receives exactly two things: the shared context above the branch point, and the messages along that branch. Sibling branches and abandoned experiments are simply not part of the request — there is no “ignore the above,” because there is no “above” to ignore.
This makes costs predictable too: you pay for the root plus one branch, not for a ten-thousand-word history you no longer care about. Why every message resends everything →
The four branches everyone ends up using
- The rabbit hole branch — environment fights and detours stay where they happened. The fix comes back to the main thread; the fight gets pruned.
- The bake-off branch — the same question to ChatGPT, Claude, and Gemini at once. See model comparison.
- The idea fork — queues vs webhooks, retainer vs project pricing. Explore both properly, prune the loser.
- The project thread — client work, admin, coding, life admin. Each in its own context. See threads.
Questions, answered
Can branches see each other's context?
Sibling branches share only the context above their common branch point — typically your files and original question. Everything below that point is private to each branch. There is no cross-contamination between threads.
How is a branch different from starting a new chat?
A new chat starts empty: you re-upload the files and re-explain the constraints. A branch starts from a chosen point, so it keeps the files, decisions, and context that matter — while staying isolated from the tangents that don’t.
Can I branch from any message?
Yes. Any message on the canvas can start a new branch, and that branch inherits everything above it — the shared root plus the thread down to that message. Branch early, branch often: branches are cheap, polluted context is expensive.
When is branching overkill?
For quick, single-topic questions — a regex, a one-off explanation — a linear exchange is fine, and every canvas also reads as a normal chat. Branching pays off when conversations get long, tasks run in parallel, or the cost of a wrong answer is high.
Try branching on the canvas
Free to start, with $5 of welcome credit. No subscription — cancel nothing, ever.