Security

Confirm-first capture

A skill is saved only after a human yes, into the author's personal workspace. It reaches the team only on evidence: reused at least three times with a success rate of 70% or more, a passing evaluation and a clean privacy scan. Org-wide promotion needs a human yes, and approvers see the full artifact, trigger, steps, checks and rationale before confirming.

Scoped, revocable keys

Personal keys (cog_me_*) are person-scoped. Team keys (cog_live_*) are org-scoped and belong in a secrets manager, never in code. A personal key cannot write team-wide approved skills without explicit promotion. All keys are revocable from the dashboard at any time.

Team boundaries

Skills, circuits, and decisions are org-bounded. A personal install cannot read another org's team brain. Org admins control who can read, write, and approve skills. A person verified as a member of several orgs can read those orgs' team and org skills; writes stay in their home org, and another member's personal memory is never included.

What we store

Approved skill artifacts (trigger, steps, checks, author, timestamps), outcome receipts (which skill, which agent, whether it helped), and agent routing signals for future retrieval improvement. We do not store raw source code, unapproved skill drafts, secrets, credentials, or communication outside the agent session. Session telemetry, such as tool output, is stored only if you turn it on, and is redacted first.

Data and deletion

Any memory can be suppressed so agents stop using it, at any time. Team skills can be archived or removed by org admins. Outcome receipts can be cleared on request. Email founders@lamlab.ai for full data deletion.

MCP transport

Cognition uses Streamable HTTP MCP, no persistent socket, no background process after the session ends. Each call is authenticated with a Bearer token scoped to the requesting user or org. Cognition does not execute code, modify files, or take actions outside returning memory and routing guidance.