Persistent memory

DeepMemory loads a string into the initial system context and exposes memory tools to the agent. It is for durable facts, preferences, or compact working knowledge—not for complete transcripts or unrestricted file storage.

var memory = new FileDeepMemory(memoryRoot);
var agent = DeepAgent.builder(model)
        .memory(memory, "customer-42")
        .build();

The namespace separates independent memory scopes. FileDeepMemory stores each namespace in a local Markdown file and accepts only letters, numbers, dots, underscores, and hyphens in its name. Its writes are atomic, making it a practical option for a local deployment with a deliberately selected private directory.

The agent receives loaded memory before the user request and may use write_memory, which is approval-protected by default. For shared, encrypted, audited, or multi-tenant deployments, implement DeepMemory against the application’s own persistence and authorization layer.

Keep memory concise. It consumes model context on every run; retain authoritative or large source material in a workspace, MCP resource, or application retrieval system instead.