Skip to main content
The Schema Call Protocol (SCP) is the core abstraction that makes Loom work. It treats agent memory as a self-describing schema — like a file system for cognition.

​
The Problem

Long-running AI agents struggle with memory:
  • Full conversation history — context window overflow
  • Unstructured summaries — imprecise retrieval, lost details
  • Vector databases — no structured schema, hard to update specific facts

​
The Solution

Loom organizes memory into typed, hierarchical fields with descriptions:
Each field carries its own description. The CM agent uses 4 standard operations to interact with memory:

​
Operations

​
read
Read data at specified schema paths. The CM agent uses this to retrieve relevant memory before generating a response.
​
write
Overwrite or append to an existing field. Used to update facts when new information is learned.
​
write
Create a new field at any path. Intermediate paths are auto-created. Used when the agent encounters information that doesn’t fit existing fields.
​
write
Create a new empty schema domain. Used to organize memory into logical groups.

​
Schema Domains

Memory is organized into domains — top-level namespaces like user_profile, medical_record, or project_context. Each domain contains a tree of fields.

​
How the CM Agent Uses SCP

During a build() call, the CM agent:
  1. Reads the user input
  2. Inspects the current schema structure
  3. Decides which fields to update or create
  4. Calls the appropriate SCP operations
  5. Repeats until no more updates are needed
The agent is LLM-powered, meaning it uses frontier model reasoning to determine what information to extract and where to store it — capturing nuanced signals that rule-based systems miss.

​
Two Schema Modes

The CM agent creates schema fields dynamically during build(). No pre-defined structure needed — the agent decides what fields to create based on the input.This is the default behavior and works well for general-purpose memory.