Agent Memory

Agent Memory provides a virtual filesystem where agents store and retrieve persistent knowledge. Each agent instance has its own structured memory that persists across conversations and executions.

How Agent Memory Works

Each agent instance has a dedicated memory store organized as a virtual filesystem. Memory files contain knowledge, configuration, conversation history, and skills that the agent references during execution.

File Types

Memory files are categorized by type, each serving a specific purpose:

Managing Memory Files

Listing Files

View all memory files for an agent instance, optionally filtered by file type or path prefix.

Reading Files

Read the content of any memory file. Files are stored as plain text (typically Markdown or JSON).

Writing Files

Create new files or update existing ones. Each write is validated before being saved:

• File path and content are checked for format correctness
• Files can be toggled on or off with the isEnabled flag
• An optional sourceSkillId tracks which skill generated the file

Deleting Files

Remove a memory file permanently. This cannot be undone.

Searching Files

Search across memory file content with full-text matching. Results can be filtered by file type and are ranked by relevance.

Initializing Memory from Templates

When you create a new agent instance from a template, you can initialize its memory with a predefined set of files.

If memory files already exist, initialization is blocked unless you explicitly force re-initialization.

Human-in-the-Loop Editing

Agents can propose changes to their own memory files. These changes go through an approval workflow before being applied.

Edit Workflow

Reviewing Edits

1. View the list of pending edits for an agent instance
2. Each edit shows the proposed changes to a specific file
3. Approve the edit to apply it to the memory file
4. Reject the edit to discard it without changes

Only the file owner (or organization admin in org context) can approve or reject edits.

Context Loading

When an agent executes, Fabric loads all enabled memory files into the agent's context. The loadContext operation gathers:

• All enabled memory files for the agent instance
• Files ordered by type priority (AGENTS_MD first, then MCP_JSON, SKILL, KNOWLEDGE, etc.)
• Content assembled into a structured context that the agent can reference

This ensures the agent has access to its full knowledge base at execution time.

Import and Export

Exporting Memory

Export all memory files for an agent instance as a JSON object. The export includes:

• File paths as keys
• File content as values
• Export timestamp

Use this to back up an agent's memory or transfer it to another instance.

Importing Memory

Import memory files from a JSON object (path-to-content mapping):

• Validation — Files are validated before import (can be skipped with validateFirst: false)
• Overwrite — Existing files at the same path are overwritten
• Batch — All files are imported in a single operation

Multi-Tenant Support

Agent memory is strictly isolated by tenant context:

• Personal (/app/agents) — Memory files scoped to your personal agent instances
• Organization (/app/{org}/agents) — Memory files scoped to organization agent instances
• Memory files from one context are never accessible in another
• Searches are scoped to the current tenant context
• Import/export operations respect tenant boundaries

---

Next Steps