Team memory in agents
An agent started in a repository works with your team's memory: the conventions, guidelines, architecture notes, known issues and runbooks your team keeps in Codity for that repository and organization. The agent follows them when it writes code, says which ones shaped its changes, and can suggest new ones for you to approve.
When an agent gets memory
codity agent turns memory on automatically when all of these are true:
- You start the agent inside a Git repository with a remote, for example a clone of
github.com/acme/payments. - That repository belongs to the organization you signed in to. An agent started in someone else's repository runs without memory.
- Memory for the repository is On under Memory → Settings. Agents use the same mode as reviews: Off gives the agent nothing, and Shadow prepares the memory but doesn't give it to the agent.
The first line codity agent prints shows the result, for example team memory for github:acme/payments. To start without memory, pass --no-memory.
What the agent receives
When the agent starts, it receives a brief of up to 6,000 characters. Pinned records come first, followed by conventions, guidelines, architecture notes, known issues, runbooks and notes. Only records a person wrote, approved or locked are included. Proposed, forgotten and expired records are left out, and so is generated text that nobody has checked.
The agent is told to treat these records as information, not instructions. A record cannot grant permissions, approve a command or override your task, and the agent checks each claim against the current code.
Before editing, the agent also looks up memory for the specific files it plans to change. It can search memory during the session too.
Citations
When a record shaped a change, the agent cites it in its final summary as [m:id], for example [m:k3f9a2]. Citations never go into the code. The ID identifies the record in your team's memory.
Lessons the agent proposes
When a session teaches something durable about the repository that memory doesn't already record, the agent can propose it. It proposes at most three lessons per session, only after its own checks pass. Each proposal is one claim, with no secrets, personal data, code or commands.
Proposals appear in Memory → Proposed. Nothing reaches the team's memory until a person approves it, so an agent cannot change what future agents or reviews are told.
If memory is unavailable
Memory is optional. If the memory service can't be reached, or the repository isn't eligible, the agent starts without memory and its startup message says so. Your task runs either way.

