Skip to main content
Replicas learns from completed work so you spend less time repeating preferences, explaining past decisions, and correcting the same mistakes. Long-term memory is available to every organization and carries context across workspaces and coding agents.

​
What gets remembered

Memory keeps information that can help with future tasks: your preferred way of working, repository conventions, technical decisions, procedures that worked, and fixes for recurring failures. It is selective, rather than a copy of every conversation. Personal memory is associated with your seat in the organization. Chats involving multiple people contribute organization memory without being assigned to one person’s memory. Organization summaries are filtered to exclude personal context and sensitive information; credentials and secrets are excluded from both scopes.

​
How memory builds

1

Work with your agent

Give feedback and explain decisions as you work. For example: “Use the existing pagination component, and verify this flow in the browser before opening a PR.”
2

Let the workspace sleep

When a workspace with memory retention enabled sleeps, Replicas captures eligible chats and extracts useful lessons. Background processing combines these summaries into personal and organization memory.
3

Start the next task

Replicas loads the latest published memory when a workspace starts or wakes. Agents search for relevant context before substantive work and can consult the supporting summaries for detail.
New memories currently come from Claude Code and Codex chats. Forked chats do not contribute new memories. Other coding agents can recall existing memory. Processing runs in the background, so a workspace that wakes immediately may receive the previous version.
Memory processing uses your connected Claude Code or Codex credentials. Keep the relevant subscription or API key connected so new memories can be processed.

​
Control what contributes

Recalling existing memory and retaining new work are separate controls. For an automation, open Advanced settings → Memory and enable Retain workspace memory to include its runs in future memory. Through the API, set config.memoryRetentionEnabled when creating a workspace. Turning retention off prevents that workspace’s work from contributing new memories. It can still recall existing organization and personal memory. The organization-wide Save chat and Canvas history setting also applies: when history saving is disabled, Replicas does not persist new chats for memory or load saved memory into workspaces. See Data retention.

​
Inspect and use memory

Ask your agent to explain the context it found:
You can also read the files in the workspace terminal:
Each scope includes a rollout_summaries/ directory with supporting summaries. These files are generated and installed as read-only context. Give corrections in chat so future memory can reflect them; editing a local copy does not update shared memory.
Memory is historical context and can be incomplete or outdated. Current instructions take priority. Keep requirements that must always apply in your repository’s agent instructions.