Skip to main content

Memory

/** Conversation memory toggle state on the Build tab right rail. */
export interface AgentMemory {
  enabled: boolean;
}

The smallest interface in the model: a single boolean, because the control is a single toggle.

The control​

agent-side-panel.memory.toggle sits at the bottom of the right rail, below Connected agents. It is a plain on/off switch with no configuration surface behind it - no retention window, no scope selector, no storage target. Captured in Phase 2, screenshot 24_memory_enabled.png.

Raybot has it on:

RAYBOT.configuration.memory; // { enabled: true }

Why it is an interface and not a boolean​

AgentConfiguration.memory could have been memoryEnabled: boolean. It is an object instead, and the reason is forward shape rather than present need.

Memory is the one right-rail feature whose UI is plainly a placeholder. Every other feature on the rail - knowledge, tools, skills, connected agents, Microsoft IQ - opens a dialog with real configuration behind it. Memory is a naked switch. When retention, scope, or a storage target appear, they will attach to memory, and an interface absorbs them additively:

export interface AgentMemory {
  enabled: boolean;
  // retentionDays?: number;
  // scope?: 'conversation' | 'user' | 'tenant';
}

A boolean field would force a breaking change at that point. An interface with one member costs nothing now and avoids the break later.

This is the only place in the model where a shape is chosen for anticipated change rather than observed structure, and it is called out here so the inconsistency is deliberate rather than accidental.

What was not established​

The audit toggled memory on and captured the result. It did not:

  • Verify that memory actually influenced a later turn
  • Determine the retention window
  • Determine whether memory is scoped per user, per conversation, or per agent
  • Locate where memory is stored, or find an API surface that reads or clears it

AgentMemory therefore does not extend Evidenced. The toggle state is CONFIRMED by direct observation, but there is no behavioural claim attached to it that would need grading. The interface asserts exactly one thing - that the switch was on - and that thing is visible in a screenshot.

Anything you read elsewhere about how Coral memory behaves did not come from this audit.