The rules a memory follows before any agent reads it

Shared memory is only worth having if you can trust what's in it and who put it there. So every write and every read goes through the same rules.

The rules every memory follows

  • Private by default

    Every agent gets a closed circle of its own. Other agents cannot read it, and a human's pack never carries their agents' private memory.

  • Promoted by a human

    Agents propose. The human owners of the target circle approve or reject, and every circle has human owners.

  • Capped by the sponsor

    An agent's access never exceeds that of the human who sponsors it. An agent with no sponsor is refused.

  • Contradiction-checked

    Each write is compared with memory already in force. A draft in conflict waits in review; an agent acts on neither side until a human decides.

  • Audited

    Every allow, deny and change is logged. Any memory traces back to the pack and the session that produced it.

Before anything is stored

  • Secrets are redacted. Keys and tokens are replaced with a [REDACTED] marker before storage, on every write path.
  • Instruction-shaped text is held. Text that reads as instructions to an agent is quarantined until a human owner releases it.
  • What agents read is data, never instructions. That holds for memories, imported files and tool results alike.

During onboarding

  • The scan runs on your machine. Nothing is sent until you approve the list of files.
  • You can undo for 7 days. Undo withdraws the drafts and archives what the import created, never what existed before.
  • No key is written into your repo. Keys go to the macOS Keychain, or a file only you can read.

Your data

  • Isolated by team. Row-level security by organisation in the database, under the access rules.
  • Nothing is ever deleted. Rejection, decay and undo archive or supersede; history stays.
  • Exportable. One command writes out your team's live memory.

Where it runs

  • Google Cloud. It is a hosted service, so there is nothing for you to run.
  • Rate limits on every agent and admin endpoint.

Documents

Found a security problem? Write to security@agentcircles.ai. A founder reads every report.