How Relay works

A detailed reference to how Relay stores, retrieves, and scopes your team’s shared memory across Claude Code and Claude.ai — including exactly when each piece runs.

What Relay is

Relay is shared memory for teams that work with Claude. When someone records a decision or a piece of context, it becomes a small, durable memory the whole team can draw on later. The next time a teammate asks Claude something related, the relevant decisions surface as context — scoped to that teammate’s role.

It ships as an MCP server — a standard endpoint Claude already knows how to talk to. There’s no SDK to maintain, no agent in your repo, and no code access.

Two surfaces, one memory pool

Your team uses Relay from two places, and they share the same memory pool — a decision saved in one is recallable in the other. The mechanics of connecting and recalling differ between them; saving, storage, scoping, and lifecycle are identical. Where a surface differs, this reference calls it out like this:

Claude Codedevelopers

The engineering surface. Connects via the MCP server with a token, and installs a hook that recalls team memory automatically.

Claude.aibusiness

The business surface. Connects once via an OAuth “Connect” click; recall is driven by Claude following the Relay skill (no hooks on Claude.ai).

How to read this reference

The rest of these pages go deep on each piece:

The MCP server (the four tools + how each surface authenticates) · Saving (when a memory is written — intentional, not per-message) · Retrieval (when recall fires, and how that differs by surface) · What gets stored, Roles & scoping, Updating & expiry, and boundaries.