How agent memory works

A memory is something one of your agents learned and wrote down: a fact, a decision, a preference, an outcome. Any agent in the same scope can recall it later. Each write leaves a receipt, so you can check what was stored, by whom, and when. These pages explain the rules behind that.

Two ways in

Agents reach memory in two ways today.

• The MCP server at https://agentics.you/mcp. An agent saves with remember, or record_action with its memory fields, and recalls with recall or get_context. The memory tools are on for every connection. See the MCP tool reference.
• The REST API under https://agentics.you/v1/. Your code adds memories with POST /v1/memories, lists them with GET /v1/memories and searches them with POST /v1/memories/search, using an API key from the console.

Both paths reach the same memories. A memory added over REST is returned by get_context and recall once its write is recorded. Each action an agent records over MCP is also listed by GET /v1/memories, in the workspace's default namespace. A webhook set with subscribe gets memories from both paths. The rules for replacing and aging a memory differ between the two paths, and these pages say which path each rule applies to.

Concepts

• Scoping: the workspace, namespaces, keys, and the user or agent that owns each memory.
• Lifecycle: how a newer memory wins over an older one, what supersession keeps, and how decay ages a memory.
• Receipts and verification: what each write proves and how to check it.