Skip to main content
An Agent is a configurable AI agent definition in MemoryLake. Each agent has a model, system prompt, capabilities, and execution policies. When you create an agent, MemoryLake automatically generates an Actor identity for it — because actors are the subject of memory, and agents need to accumulate knowledge just like human users do.

​
Key Concepts

​
Versioning

Agent configuration is versioned. When you change the model, system prompt, or policies, you create a new version rather than modifying the existing one. This gives you a history of configurations and the ability to roll back. Identity fields (name, description, metadata) are updated in place via Update Agent — they don’t create a new version.

​
Workspace Binding

Agents are created globally, then bound to specific workspaces. An agent can only operate within workspaces it is bound to.

​
Agent Memory

An agent accumulates facts of its own — what it has learned, and how it is meant to behave. They live on the agent, read by its agent id within a workspace:
You can list, read, add, update and forget them, read their change history, and review and resolve conflicts between them — see Agent Facts and Agent Memory Conflicts. Address them by agent id: the actor endpoints do not accept an agent. For how agent facts sit alongside the other scopes, see Memories.

​
Forking

Fork an agent to copy its latest version into a new, independent agent with its own actor. The copy starts at version 1 and does not take the source agent’s version history, metadata, workspace bindings or memory.

​
A2A Protocol

MemoryLake agents are accessible via the standard A2A (Agent-to-Agent) protocol. Use any A2A-compatible client to send messages to an agent, with MemoryLake’s metadata.memorylake extension providing memory context (actor identity, project bindings, model overrides).

​
Endpoints