The AI That Talks Back to Its Own Memories

Most memory systems are read-only from the agent's perspective. The agent can retrieve memories, but it cannot meaningfully critique or update them. The Cognitive Substrate is designed to change that.

Note: Memory Critique Events are a planned extension to the current architecture. The foundation is in place - agents already carry a critique field on their result type - but the full MCE pipeline described below is forward design, not yet implemented.

The Seed: Agent Critiques Today

The current AgentResult type already includes a critique field - agents in the multi-agent reasoning loop can annotate their own proposals with a critique string. This is the seed:

// packages/core-types/src/agent.ts

export interface AgentResult {
  readonly agentId: string
  readonly agentType: AgentType
  readonly proposal: string
  readonly reasoning?: string
  readonly confidence: number
  readonly riskScore: number
  readonly retrievedMemories: ReadonlyArray<string>
  readonly critique?: string // the foundation MCEs will build on
  readonly score?: number
}

The next step is promoting that critique string into a first-class structured event that flows back into the memory pipeline.

Memory Critique Events

The planned Memory Critique Events (MCEs) extend this foundation, allowing the agent to actively challenge its own knowledge base with structured feedback.

When an agent retrieves memories through the arbitration layer, it can now emit structured critiques containing:

• memory_id or cluster_centroid - identifying what is being challenged
• critique_type - one of: contradiction, outdated, overgeneralized, missing_context, high_confidence_error
• confidence_score - how certain the agent is about this critique
• explanation_embedding - optional semantic representation of the reasoning
• suggested_replacement - for high-confidence critiques, a proposed alternative

These events are published to a dedicated memory-feedback Kafka topic and consumed by a specialized Memory Critique Worker.

What the Worker Can Do

The worker can trigger several downstream actions:

• Downgrade the trust score of the specific memory or the entire cluster it belongs to
• Trigger early re-consolidation of the affected abstraction ladder branch
• Inject a new competing memory at the same abstraction level
• Raise the suppression threshold for similar future retrievals

The Power of High-Level Critiques

Particularly significant is the ability to critique higher-level abstractions. A critique targeting a Concept or Principle level memory can cascade through the compression ladder, potentially invalidating dozens of lower-level memories derived from the now-challenged abstraction - all in a single action.

This creates a self-correcting, adversarial memory system where the agent is no longer a passive consumer of its own memories. It becomes an active participant in shaping what it knows.

Over time, this feedback loop allows the system to develop something resembling epistemic humility - the ability to recognize when its own understanding is flawed or incomplete. It also creates the foundation for true continual learning, where knowledge is not just accumulated but actively refined and debated by the agent itself.