Nous AI Agent Framework Introduces Minsky's Society of Mind

What changed

The Nous AI agent framework, developed by tfatykhov, has been released with version 1.0.0. This framework is built upon Marvin Minsky's "Society of Mind" principles and decision intelligence concepts, aiming to create AI agents that can think, learn, and grow, rather than merely acting as stateless reactors.

Unlike conventional AI agents that forget after each interaction or rely solely on text-based memory retrieval, Nous introduces a structured memory system. This architecture is designed to mirror cognitive processes, enabling agents to learn from past decisions, calibrate their confidence levels, and exhibit proactive autonomy. The framework incorporates several core concepts from Minsky's work, including K-Lines for context bundles, Censors for guardrails, and Frames for interpreting input. It also integrates principles from "Cognition Engines," such as a Decision Memory, Pre-Action Protocol, Deliberation Traces, and Calibration mechanisms.

The operational cycle of a Nous agent follows a detailed loop: SENSE (stimulus reception), FRAME (interpretation), RECALL (hybrid memory search using K-lines), DELIBERATE (pre-action protocol involving decision memory queries and confidence assessment), ACT (execution), MONITOR (self-assessment by a B-brain), and LEARN (memory update including decisions, K-lines, calibration, and guardrails).

Memory in Nous is architected across different layers: Slow (Identity), Medium (Knowledge - Facts, K-Lines, Episodes), Fast (Working Memory, Events), and Persistent (Intelligence - Decisions, Calibration). This memory is stored in PostgreSQL with pgvector across three schemas: brain, heart, and nous_system, allowing for multi-agent hosting within a single database.

Nous agents are designed to grow through "administrative growth," focusing on improving management of existing knowledge rather than simply accumulating more. This involves building detours, using friction to guide behavior, and implementing censors to block failures. The current growth level is at Level 3 (learning from outcomes via calibration), with Levels 4 (monitoring thinking) and 5 (improving processes) planned, requiring a fully autonomous B-Brain.

Confidence and calibration are central to Nous. Agents record confidence scores for decisions, which are then compared against outcomes to measure calibration accuracy using Brier scores. A write-time calibration scale adjusts recorded confidence towards observed accuracy.

Why it matters for builders

Nous provides a foundational framework for developers looking to build more sophisticated AI agents. Its emphasis on structured memory, decision learning, and self-monitoring offers a path beyond current agent limitations. Builders can leverage this architecture to create agents that exhibit more robust reasoning, adapt to new situations based on past experiences, and operate with greater autonomy.

Practical impact

Developers can explore the Nous framework by deploying it from scratch using the provided Quickstart Guide. Experimenting with the agent's operational loop and memory architecture can offer insights into building agents with enhanced cognitive capabilities. The framework's implementation of Minsky's principles and decision intelligence provides a unique opportunity to test and develop advanced agent behaviors.

Caveats and source limits

The source material indicates that some Minsky concepts, such as Polynemes, Nemes, Pronomes, and Attachment Learning, are still planned. Similarly, Levels 4 and 5 of agent growth, involving self-monitoring and process improvement via a B-Brain, are in development. The specific empirical data for the calibration scale's original factor and its subsequent retirement are mentioned but lack detailed context on the refit process. The source is a GitHub repository, and while it mentions a "fresh release," specific release notes or changelogs beyond version 1.0.0 are not detailed.