Relay

Saving memory

When does Relay actually write a memory? On Claude Code, an automatic mode (the default) saves notable new facts on its own; you can also always save explicitly. Either way, only durable facts are stored — never a per-message transcript.

Saving explicitly

Whatever the save mode, you can always tell Claude to save — these explicit triggers work on both Claude Code and Claude.ai:

Claude proposes, you confirm

When you make a durable, team-relevant statement (a decision, a convention, client context, a correction), Claude proposes saving it and asks before writing. This keeps the pool signal, not chatter.

You ask directly

Say “share this with the team” or “Save this to our Relay team memory: …” and Claude saves in one shot — no confirmation needed. This is the explicit, one-shot path.

Only what matters

Casual back-and-forth, questions, and throwaway context are not stored — only durable facts a teammate would benefit from later, whether saved manually or by automatic mode.

Rule of thumb: only durable, team-relevant facts get saved — decisions, conventions, client/product context — never casual chatter. Reading from memory is constant and automatic (see Retrieval); writing is selective.

The two trigger shapes

Any of these route to team_memory_add. The first asks before saving; the explicit forms save immediately:

saving — what to type
# Inferred — Claude proposes the entry, you confirm:
"We decided to use Postgres over Mongo for this project."

# Explicit one-shot — saves immediately, no confirmation:
"Share this with the team via Relay: Acme wants invoices monthly, not per-seat."
"Save this to our Relay team memory: the staging URL is staging.acme.test"

When you save, you (or Claude) pick an area (e.g. product, sales, tech, ops) and an audience (all, dev, or commercial) — these drive scoping and retrieval. See What gets stored.

Automatic save

When you install Relay in Claude Code you choose a save mode — Automatic (the default) or Manual. In automatic mode Claude saves notable new facts on its own; in manual mode nothing is saved unless you use the explicit triggers above. It is deliberately conservative: it checks only every several turns, and only when a turn produced a genuinely new, durable, team-relevant fact — never every message, never a transcript.

Claude Codedevelopers
Automatic mode installs a Stop hookthat runs after each response. About every 7 turns, and only on meaningful turns, it injects an instruction asking Claude to save any genuinely new durable facts. The hook never writes memory itself — Claude decides what's worth saving with full context. Switch modes by re-running the install script.
Claude.aibusiness
No hooks exist on Claude.ai, so automatic save is best-effort and skill-driven — Claude may offer to save notable facts as they come up, but there is no mechanical trigger. For reliable saves on Claude.ai, use the explicit triggers above.

Either mode only ever stores durable, team-relevant facts — never trivial chatter or a transcript.