Team Memory
A shared, human-reviewable memory for your workspace: AI agents connected over MCP save durable facts (decisions, conventions, gotchas) as notes, and every future AI session - anyone's - starts already knowing them.
Before you start
- The workspace OWNER must be on the Pro plan or higher (Pro: 200 memories per workspace, Max: 500). On the free plan the feature is locked.
- An AI agent must be connected over MCP - see the Claude Code installation guide. Memories are written and read through that connection. Claude Code installation
- The feature is ON by default for eligible workspaces. Nothing to install.
Turning it on / off
Workspace Settings -> AI tab. Only the workspace owner can flip the Team Memory switch; the same tab shows a live 'Memories used N / cap' counter.
Turned off: agents' save_memory calls are politely rejected, the review strip and sidebar badges disappear, and existing memory notes are kept read-only as ordinary notes. Every other MCP tool (docs, reports, reading notes) keeps working - turn it off freely if the workspace only uses MCP for docs or test reports.
Plan limits
Team Memory is a Pro-and-up feature, keyed on the workspace owner's plan. The cap is per workspace; updating an existing memory never counts against it.
| Free | Pro | Max | |
|---|---|---|---|
| Memories per workspace | 200 | 500 |
How it works day to day
- 1While working, an agent saves a durable fact (a human corrected it, the team settled a decision, a trap cost real time) with the save_memory tool. Each memory becomes a note in the Memory session, filed by type.
- 2An index note named MEMORY.md keeps one line per memory. At the start of every MCP session the server announces it, and the agent reads the index before its first task - opening an individual memory only when relevant, to keep token use low.
- 3Saving the same memory id again UPDATES that note (the old body goes to version history) instead of creating a duplicate.
- 4Secrets are rejected on write, and one memory is capped at 4,000 characters - long material belongs in a normal note with a reference memory pointing at it.
Reviewing what agents saved
New agent-saved memories show an amber strip at the top of the workspace until someone on the team has looked. Open one to read it, tick it to acknowledge, or use Mark all reviewed.
A memory is a normal note: edit or delete it and your version wins - agents always read the current note content. If an agent later updates a memory a human already reviewed, it returns to the review strip.
Saving a memory yourself
Select text in any note and press the brain button in the formatting toolbar. Pick a type (decision / convention / gotcha / feedback / reference) and save - human-saved memories skip the review strip.
Exporting to a repo file
Ask your agent to 'export team memory as CLAUDE.md' - the export_memories tool compiles every memory into one markdown document the agent can write into your repo. The notes in PugBase stay the live source of truth.