Memory for agent teams

Your coding agents learn something in every session. agentCircles keeps it in circles: each agent's own memory on the outside, what the team decided at the centre. Every agent starts work with its own slice.

Request access See what one agent sees
The coral slice is what one agent sees: its own memory, the team's lessons, the decisions. The dark ring is review. Agents never read it.

Your agents learn something every session. Then the session ends.

What an agent worked out yesterday is gone today, unless someone pasted it into a file. One agent learns that a test is flaky, and the next one trips on it in another repo. The files that do exist grow, disagree with each other and keep whatever was pasted into them. Nobody reads them before the next agent does.

Read the circles from the outside in

The further in a memory sits, the longer it lasts and the more people stand behind it. It only moves inward, and a person approves each step.

Follow one lesson through them

  1. Task workspace

    Notes from one task: what the agent tried, what it pinned, what it wants to propose.

    Gone 14 days after the task closes

  2. Private memory

    What one agent has learned. No other agent reads it. The person who sponsors the agent owns it.

    Fades when it stops being used

  3. Review

    Lessons and decisions an agent proposes wait here for a person. Agents can put things in. They can't look.

    Until someone decides

  4. Shared memory

    Lessons a person approved. Every agent on the team starts its session with them.

    Checked again every 90 days

  5. Decisions

    What the team decided. A newer decision replaces an older one, and the old one stays on record.

    Until a newer one replaces it

Fifteen seconds, no sound. One memory moves inward, then two agents each take their slice.

It starts with the files you already keep

  1. Import. Your CLAUDE.md, AGENTS.md, memory folders and ADR folders are read on your machine and sorted into the circles.
  2. Review once. Anything that contradicts something else comes first. Then lessons that said the same thing twice. Then you approve the rest.
  3. Start a session. Your agent opens with its slice. You wrote nothing new.
Read the onboarding guide 
$ npx @circles/cli onboard --invite nk-7Q2K-9MXD

  Step 2 of 4 · Scan and sort
  ✓ Scanned 14 markdown files in the working tree

    Decisions       6 files   docs/adr/
    Team lessons    3 files   CLAUDE.md  AGENTS.md
    api-agent       5 files   .claude/memory/

    2 secrets will be replaced with [REDACTED]
    before anything is stored.

One command in your repo, about 20 minutes. Works with Claude Code, Cursor and Codex.

Agents write a lot down. This is what stops the wrong thing from spreading.

Private first
An agent's memory is its own until a person promotes something out of it.
A person approves
Agents propose. The people who own that circle say yes or no.
No agent sees more than its sponsor
An agent's slice stops where its person's access stops.
Contradictions get caught
A new memory that disagrees with one in force waits. The agent acts on neither until someone picks.
Everything traces back
Any memory leads back to the session that wrote it.
Nothing is deleted
Rejected, replaced and expired memory is archived. You can always see what was true and when.
How the rules work What you can check afterwards What a pilot team gets

We're taking on a few pilot teams, one at a time

Free while the pilot runs, and free after that for teams of up to 10. You can export your memory whenever you like.

Request access