Decisions are the context that matters
by Dev Ashish · 7 October 2026 · 2 min read
Most of what you type to an AI tool is disposable. What the next tool actually needs is the residue of decisions, the constraints you set and the approaches you already ruled out, each with a note of where it came from. Relay captures that structure, keeps it on your machine, and hands only the relevant part to whichever tool you open next.
Every new session with an AI tool starts blank. You paste the context back in, re-explain the decisions, and rebuild the thread, every time, in every tool. Relay exists to end that. The interesting question is what, exactly, it should carry.
Not everything you typed
The obvious answer is to save every conversation and search it later. It works badly. Most of a working session is exploration: options tried, half-formed ideas, a wrong turn and its correction. Feed all of that into the next session and the few lines that matter are buried, recall gets noisy, and the context window fills with things the model should ignore.
The residue of decisions
What a new session actually needs is smaller and more specific:
- Decisions: we chose Postgres, not Mongo.
- Constraints: no new dependencies in this package; the API must stay backwards compatible.
- Rejected approaches: we tried caching at the edge; it broke auth, so not that.
- Open questions: still unclear how retries should behave.
The rejected approaches are the most underrated. A new agent that doesn't know an idea was already tried and abandoned will cheerfully propose it again.
Relay captures these as typed units rather than text, so a tool can ask for "the constraints on this module" and get exactly that.
Every fact keeps its source
A memory you can't check is a memory you can't trust. Each unit in Relay records where it came from, which conversation or which file, and you can correct it. When a new session obeys a constraint it was never told about in that session, you can see why.
On your machine
Relay runs locally, and tools read it through an MCP server, so Claude, coding agents and other tools all draw on the same context without that context living inside any one vendor's product. Your thinking stays yours and moves with you between tools.
It is the same idea we run the company on. Our product hub is a store of decisions, constraints and what we got wrong, and every agent starts from it. You can read about that in how we work.
Questions
What is a context layer for AI tools?
A store of what you have decided and learned while working with AI, kept outside any one tool, so that every new session in any tool can start from it.
Why not just save the whole chat history?
Most of a chat is exploration. Feeding all of it back makes recall noisy and expensive, and buries the few lines that matter. Structure, decisions and constraints with their sources, is smaller and more reliable.
Where does Relay keep the data?
On your own machine. Tools read it through a local MCP server, so your context is not deposited in any one vendor's memory.