Project memory vs personal memory: how Relay decides what goes where

Relay keeps two kinds of memory and they behave differently on purpose. Mixing them up is how memory products become junk drawers, so the separation is strict.

Project memory: decisions with receipts

A project holds the durable truth of one piece of work: the objective, decisions (with why), constraints, open tasks, shipped artifacts. It's built from your captured chats and agent sessions, and it's what the brief is generated from. When a decision gets superseded, the old one is archived, not silently overwritten — you can always trace why something is the way it is.

Personal memory: facts about you

Personal memory holds things that outlive any project: you prefer TypeScript, you're based in Almaty, you're launching in March, your friend's startup does logistics. It's auto-categorized folk-style (person, company, event, note) and follows you across every tool.

How routing decides

When Relay captures a chat, it scores which project the conversation belongs to: name mentions, overlap with the project's description and saved context, previous associations for that conversation. A few rules keep it honest:

• Your manual pick always wins. If you set a project for a chat, Relay won't second-guess it.
• A project name being mentioned is not enough to silently reroute a save. If Relay isn't sure, it asks instead of guessing.
• Personal is never auto-selected as a dumping ground. Full chats only go there when you deliberately park a conversation on personal.
• "What do you know about me" style chats route to personal, even if your project gets name-dropped in them.

Steering it

Give your projects a real description, that's what the router matches against. Use the chip's project switcher when you start a chat that belongs somewhere unusual. And prune:/forget anything that shouldn't have been saved. Memory you trust is memory you actually use.