From Knowledge Operating System → Persistent Cognitive State

CogMem moves towards conciousness

I wanted to share an update on my ongoing PhD research, where I’m exploring cognitive memory architectures for multi-agent systems that enable persistent state, continuity, and self-improvement over time.

Over the past year, my research has undergone a meaningful shift—one that reflects a broader transformation happening across the field of AI agents.

What began as an exploration of a Knowledge Operating System (KOS) has evolved into something more ambitious:

A Persistent Cognitive State (PCS) agent architecture

This post reflects that journey—what changed, why it changed, and the research that shaped the transition.

Phase 1: The Knowledge Operating System (KOS)

The original vision for CogMem was grounded in a simple idea:

What if AI agents had an operating system for knowledge?

The original KOS model focused on:

• Structured ingestion (documents, conversations, data streams)
• Multi-modal retrieval (vector, graph, full-text)
• Planning-driven execution pipelines
• A “memory kernel” coordinating agent behavior

Architecturally, this aligned with emerging cognitive agent frameworks where memory, action, and decision-making are first-class components.

For example, the CoALA framework (Sumers et al.) formalizes agents around:

• Working memory
• Long-term memory
• Structured action loops

This validated the KOS direction: memory is foundational.

The Problem: KOS Was Still Too Static

Despite its strengths, the KOS model had a critical limitation:

It treated memory as infrastructure… not as cognition.

Most systems in this space focus on:

• Storing information
• Retrieving information
• Augmenting prompts

But they lack something essential:

Continuity over time

At this point, it became clear:

We don’t just need better memory systems.

We need agents that operate from a persistent cognitive state, not just a temporary context window.

Phase 2: Memory Becomes Dynamic (Agentic Memory)

The next wave of research pushed this further.

Memory as structure → Memory as process

• A-MEM: memory becomes a self-organizing, evolving network
• Mem0: memory becomes efficient, selective, production-ready
• Memento / ReasoningBank: memory becomes learning itself

The trajectory is clear:

Memory is moving from storage → to adaptation → to learning

The Breakthrough Insight

Across all of this work:

Memory systems were improving… but the agent itself remained fragmented.

• Retrieval is separate from reasoning
• Memory is separate from planning
• Learning is externalized

What’s missing is a central cognitive control system.

Phase 3: The Persistent Cognitive State (PCS) Agent

I originally called this the "personalized planning agent". Which is essentially still its primary function but this is where the architecture fundamentally changes.

CogMem is no longer just a memory system.

It becomes a cognitive system.

At the center is:

A Persistent Cognitive State

🧠 Insight 1: Working Memory as the Active Core (Baddeley & Hitch, 1974)

Baddeley & Hitch reframed memory as something critical:

Not storage—but a system for holding and manipulating information during reasoning.

Their model introduces:

• A central executive (attention + control)
• A working buffer (active context)
• Integration with long-term memory

Mapping to Persistent Cognitive State:

• The agent’s active context window = working memory
• The planning agent = central executive
• Retrieval = controlled activation of long-term memory

This creates a critical distinction:

Memory is not all equal.

• Most memory is inactive
• A small subset becomes active cognition

This is exactly what current RAG systems lack.

🌐 Insight 2: Global Workspace as the Control Mechanism (Baars / Dehaene)

Global Workspace Theory explains how cognition becomes coordinated:

Multiple specialized processes compete for attention,

and what “wins” gets broadcast across the system.

Key ideas:

• Competition for attention
• A shared workspace
• Global broadcast to specialized modules

Mapping to Persistent Cognitive State:

• Specialist agents = distributed cognitive processes
• The planning agent = the workspace
• Selected memory + context = what gets “broadcast”
• Tool calls + agent coordination = execution of broadcast decisions

This is the missing piece in most agent frameworks:

They have multiple agents—but no true coordination mechanism grounded in cognition.

🧩 What PCS Actually Is

Putting it together:

The PCS agent maintains a continuously evolving internal state composed of:

• Working memory → what the agent is actively thinking about
• Autobiographical memory → what has happened over time
• Semantic memory → structured knowledge of the world
• Reasoning memory → learned strategies and outcomes

But more importantly:

It controls what enters, what stays, and what is acted upon

From Memory System → Evolving Persistent Cognitive State

Old Model (KOS) → New Model (PCS)

• Memory as storage → Memory as identity
• Retrieval-driven → Attention-driven
• Stateless agents → Persistent agents
• External memory → Integrated cognition

The Role of the Planning Agent (Reframed)

This evolution completely redefines the planning agent.

Before:

• A coordinator of tools

Now:

• A cognitive control system

It:

• Selects relevant memory (attention)
• Maintains working context (workspace)
• Coordinates specialized agents (broadcast)
• Updates memory (learning loop)

It is not just planning

It is maintaining continuity of thought over time

Why This Matters

1. Agents become longitudinal

They persist across sessions, not just prompts.

2. Learning becomes intrinsic

Memory accumulation is learning.

3. Multi-agent systems become coherent

Not just multiple agents—but a coordinated cognitive system.

Where This Is Going

This sets up the next layer:

The Continuity Engine

A system responsible for:

• Maintaining cognitive state
• Managing attention and memory flow
• Coordinating agent networks
• Driving self-improvement loops

Final Thought

The biggest realization in this journey:

The future of AI agents is not better prompts, tools, or even models.

It is:

👉 Persistent cognitive state

From:

• Stateless responses To: Continuous identity

From: Memory as context To: Memory as mind

If you’re working on agent systems, memory architectures, or cognitive AI—I’d love to hear how you’re thinking about this shift.

References / Influences

• Baddeley, A. D., & Hitch, G. (1974). Working Memory
• Baddeley, A. (2000). The Episodic Buffer: A New Component of Working Memory
• Baars, B. J. (1988). A Cognitive Theory of Consciousness
• Dehaene, S., & Changeux, J.-P. (2011). Experimental and Theoretical Approaches to Conscious Processing
• Sumers, T. R., et al. (2024). Cognitive Architectures for Language Agents (CoALA)
• Chhikara, P., et al. (2025). Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory
• Xu, W., et al. (2025). A-MEM: Agentic Memory for LLM Agents
• Zhou, H., et al. (2025). Memento: Fine-tuning LLM Agents without Fine-tuning LLMs
• Ouyang, S., et al. (2025). ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory