How do we get to AGI?

Mini series: Cognitive Architectures (3/5)

Welcome back to The AGI Observer. In Part 3 we build on our Introduction (0/5, Part 1 on Scaling Deep Learning and Part 2 on Neuro-Symbolic Systems to examine Cognitive Architectures: what they are in plain language, why they matter now, who is building them, where the bottlenecks are, and how we will know they are working. As before, we end with a sample educational virtual portfolio to show how this pathway maps to real companies and infrastructure. Educational only; not financial advice.

Nate Derbinsky - one of the main proponents of this approach

---

Executive summary

Thesis. Cognitive architectures aim to assemble an artificial mind: cooperating modules for perception, working memory, long-term memory, planning, action/tool-use, and self-monitoring that actually work together.
Why it matters. More transparent and controllable by design: users can see steps, sources, and decisions. This is a plausible route to “human-like” generality because skills can transfer across tasks.
Near term. Project agents that remember across days, plan first, execute with tools, and provide timeline-style traces of what happened.
Risk. System complexity, brittle interfaces between modules, latency from multi-step plans, and slow iteration due to weak end-to-end benchmarks.
What to watch. Agent frameworks with persistent memory, global-workspace-style attention (a shared board where ideas compete), and reliable cross-task transfer.

---

1) What “cognitive architecture” means in practice

Instead of one giant model guessing everything, roles are separated and coordinated:

• Perception understands inputs (text, images, audio, video).

• Working memory keeps the active details on hand.

• Long-term memory stores and retrieves facts and past episodes.

• Planning breaks goals into steps and chooses tools.

• Action & tools run searches, code, spreadsheets, databases, or APIs.

• Self-monitoring checks progress, fixes errors, and asks for help when needed.

---

2) Why this pathway matters

Real projects last longer than a chat window. Cognitive architectures keep momentum across days and make outcomes auditable. They also support skill transfer: once a system learns “how to reconcile a spreadsheet” or “how to draft a risk summary,” that procedure can be reused in new contexts without starting from zero.

---

3) How it works under the hood

• Structured memory, not guesswork. Items are saved with keys, who/what/when/why, so retrieval is deliberate.

• Plans before prose. The agent writes a short plan, then executes it; plans are cheap to update when reality disagrees.

• Checks at each step. Rules and validators verify numbers, formats, and policies.

• Global workspace. A central board where candidate ideas or memories are surfaced and the system selects what to focus on.

---

4) Where we are today

Agent frameworks are maturing: plan-and-execute loops, tool-calling, and memory stores are becoming standard. Long context helps, but structured memory remains more reliable and cost-effective. Early wins appear in policy and legal drafting with citations, finance and operations tasks that repeat monthly, “living” research files that update as data arrives, customer cases that span days, and software maintenance where planning and history reduce rework.

---

5) Who is building it

• SOAR (John Laird, Paul Rosenbloom)

• ACT-R (John R. Anderson)

• Global Workspace / LIDA traditions (Bernard Baars; Stanislas Dehaene on neural evidence)

• OpenCog / Hyperon (Ben Goertzel)

• Numenta (Jeff Hawkins) on cortex-inspired approaches

• Academic groups at CMU, JHU, RPI, and others

  

---

6) Sample educational virtual portfolio - Cognitive Architecture stack

Education only; not financial advice. Equal weight within buckets.

• Agent & workflow platforms (30%)
  Microsoft (MSFT); Alphabet/Google (GOOGL); ServiceNow (NOW); UiPath (PATH)

• Memory, search & data (20%)
  Snowflake (SNOW); Elastic (ESTC); MongoDB (MDB)

• Integration & observability (15%)
  Datadog (DDOG); Confluent (CFLT)

• Compute & acceleration (15%)
  NVIDIA (NVDA); AMD (AMD)

• Collaboration & content backbone (10%)
  Atlassian (TEAM); Box (BOX)

• Industrial & edge execution (10%)
  Rockwell Automation (ROK); ABB (ABB)

---

Closing thought

Part 1 showed how Scaling Deep Learning keeps expanding raw capability and lowering costs. Cognitive architectures organize that capability into goal-directed systems that remember, plan, adapt, and explain themselves, a necessary step if we want dependable intelligence on real work that stretches beyond a single prompt.

---

---

Prepared by AI Monaco

AI Monaco is a leading-edge research firm that specializes in utilizing AI-powered analytics and data-driven insights to provide clients with exceptional market intelligence. We focus on offering deep dives into key sectors like AI technologies, data analytics and innovative tech industries.

---

Disclaimer:

This report (including any accompanying materials, emails, or communications) is provided by AI Monaco and Leonardo Fabbri, both operating from Monaco, Monte Carlo, and abiding by the relevant laws and regulations of the European Union. The content herein is strictly for informational and educational purposes and does not constitute financial, investment, legal, or any other type of professional advice. It should not be interpreted as a solicitation, recommendation, endorsement, or offer to buy or sell securities, financial instruments, or investment strategies. Accuracy & Forward-Looking Statements. The opinions and analyses in this report are based on publicly available information believed to be accurate at the time of writing but are subject to change without notice. Any forward-looking statements or estimates are speculative and should not be relied upon as guarantees of future performance. AI Monaco and Leonardo Fabbri make no representations or warranties, express or implied, regarding the completeness, accuracy, reliability, suitability, or availability of the information contained herein. Non-Personalized Advice & Fiduciary Relationship. AI Monaco and Leonardo Fabbri are not registered investment advisors, brokers, or dealers and do not provide personalized investment advice or execute trades. By using this report, you acknowledge and agree that no fiduciary relationship has been established between the authors/providers and the reader. Any reliance you place on the information in this report is strictly at your own risk. Liability & Risks. Investing involves risks, including the possible loss of principal. Past performance is not indicative of future results; the value of investments and income derived from them may fluctuate, and can go down as well as up. You are strongly advised to consult with a qualified financial advisor, attorney, tax advisor, or other professionals before making any investment decisions or engaging in financial transactions, as your personal financial situation, investment objectives, risk tolerance, and tax implications have not been considered in the preparation of this material. Limitations of Liability. To the fullest extent permitted under the laws of Monaco and the European Union: No Liability for Damages: AI Monaco and Leonardo Fabbri shall not be liable for any direct, indirect, incidental, special, consequential, or exemplary damages (including, but not limited to, financial loss, lost profits, or trading losses) arising out of or in connection with the use of or reliance on this report. Data & Market Uncertainties: AI Monaco and Leonardo Fabbri disclaim liability for inaccuracies in third-party data, market changes, regulatory shifts, or any other events that may affect the validity or timeliness of the information provided. No Assumption of Responsibility: AI Monaco and Leonardo Fabbri do not assume responsibility for any actions taken by individuals or entities based on this report. Any claims, disputes, or legal actions initiated against AI Monaco or Leonardo Fabbri—including but not limited to those related to securities laws, defamation, misrepresentation, or negligence—are disclaimed to the fullest extent permitted by applicable law. No Obligation to Update. All opinions expressed herein reflect the authors’ judgment as of the publication date. AI Monaco and Leonardo Fabbri are under no obligation to update or maintain current the information contained in this report. Not an Offer or Solicitation. This report is not an offer or solicitation to buy or sell any securities or financial instruments. Readers should seek independent legal and financial counsel before engaging in investments or other financial decisions. Acceptance of Disclaimer. By accessing or using the information contained in this report, you acknowledge that you have read this disclaimer in full, understand it, and agree to be bound by its terms. If you do not agree with any part of this disclaimer, you should refrain from using the information provided here