Follow one lesson from a Tuesday afternoon to every agent
An agent finds out that your courier does not move live plants over the weekend. Watch where that goes. It takes six steps, and a person is in one of them.
Prefer to scroll through it? See the story-
1. A session starts, and the agent gets its slice
Its own memory, the team's lessons and the decisions that matter for the task. It didn't have to ask.
-
2. It works, and takes notes
An order arrived wilted because it sat in the courier's depot all weekend. That goes into the task's notes, on the outer circle.
-
3. It remembers what's worth keeping
The note becomes one of its own memories. From now on this agent knows, in any session. Nobody else's agent does yet.
-
4. It proposes the lesson to the team
The draft goes into review and stops there. The agent gets a receipt. It can't read the draft back.
-
5. A person approves it
One step. If the draft contradicts something already in force, the reviewer sees both sides first and picks.
-
6. The next agent starts with it
Another agent, another job, a session next week. The lesson is in its slice, and written back into the files your agents already read.
What happens while nobody is watching
Three jobs run every night. None of them can put anything into shared memory without you.
- Notes become memories
- Closed tasks are read back. What an agent marked, repeated or pinned becomes one of its memories. If two agents learned the same thing, one lesson draft goes to review.
- Old memory fades
- A memory no session has used for about two months starts to fade, then is archived. Task workspaces go 14 days after the task closes. Decisions never expire; they are flagged for a second look.
- Contradictions are found
- Memories that disagree are paired. A rule settles most pairs, and the rest are put in front of a person. Until someone decides, agents act on neither.
Four things it never does
- Call a model to build a slice
- Recall is a lookup. The same question gets the same slice, and it is fast.
- Let an agent approve its own lesson
- Agents propose. Only the people who own a circle let something into it.
- Treat memory as orders
- What an agent reads is data. Text that reads like instructions is held back until a person releases it.
- Delete
- Rejected, replaced and faded memory is archived. You can always see what was true and when.
What it plugs into
Your agents reach it through MCP, the standard way agents use tools. Any agent that speaks MCP can connect. Claude Code also gets its slice on its own when a session starts. Cursor and Codex connect through the same tools.
Your files stay where they are. The instruction files, memory folders and decision records your agents already use sync both ways, and approved memory is written back into them.
You use one command-line tool. There is nothing to deploy: it's a hosted service on Google Cloud.
Read the onboarding guide