Memory best practices
Memory's value compounds when teams treat it like a shared notebook, not a transcript dump. Five rules to keep it useful.
1. Save decisions, not deliberations
Good memory: "We picked Ably for realtime — Cloud Run can't hold long WebSockets."
Bad memory: a 40-message debate with three options on the table.
The agent reads memory at every turn. A clear conclusion in 2 sentences earns its keep across hundreds of future replies. A long deliberation just clutters context.
2. Pick visibility deliberately
ORG— facts the whole team needs (deploy region, on-call rota, holiday schedule).SPACE— project-specific decisions (launch date, design constraints, integration choices).PRIVATE— your personal preferences the AI should respect ("I prefer Tailwind over plain CSS"). Stays yours; teammates never see it.
Visibility is enforced at the SQL layer, not by the AI. Saving as ORG means the whole team can read it from now on. Saving as SPACE in the wrong Space is a leak you have to manually delete.
3. Use the @mention syntax to write
The cleanest way to save: ask the agent to do it.
- "@claude remember for the team that the launch date is April 30."
- "@gpt-5 save to this Space: design language is utility-first Tailwind, no Material UI."
The agent calls add_memory with the right visibility based on what you said. You can also write directly via the SDK or REST API when you want exact control.
4. Don't save anything you wouldn't paste in chat
Treat memory as semi-public. Anyone with org access can read ORG memories; anyone in a Space can read SPACE memories. Don't save credentials, customer PII, or anything covered by an NDA.
For credentials specifically — Switchy stores OAuth tokens for connected MCPs in Google Secret Manager, not in memory. Don't paste API keys into the chat composer.
5. Prune as you go
Stale memory is worse than no memory — the AI confidently cites facts that have changed. When a decision gets reversed, save the new one and delete the old. The dashboard memory list lets you delete in one click.
What this isn't
Memory is not a knowledge base or a wiki. It's the team's scratchpad of facts the AI should know. If you need versioned, hyperlinked, hierarchical content, use Notion (and connect it via @notion so the AI can search it on demand).
Next
- Concepts: memory — the visibility model in depth.
- Connect Notion — for the wiki-style content the AI should reach.