- Hyperon and PRIMUS fuse symbolic reasoning with neural learning through shared Atomspace memory to support broader intelligence aims.
- PRIMUS cognitive loops guide goals and pattern discovery while integrated tools handle tasks within one unified memory structure.
SingularityNET’s lead scientist Ben Goertzel recently released a detailed whitepaper presenting Hyperon and PRIMUS as an integrated architecture aimed at evolving current AI systems into artificial general intelligence (AGI). The project brings together symbolic reasoning and neural learning in one structure, with the goal of eventually reaching superintelligence.
SingularityNET stated,
“We present the Hyperon AGI platform and the associated PRIMUS cognitive architecture as a unified neural-symbolic framework designed to advance from current AI to AGI to beneficial superintelligence.”
Unlike many AI frameworks built from disconnected modules, Hyperon uses a unified method for processing perception, memory, and learning. It stores all system content, including goals, information, processes, and neural patterns within a structure called Atomspace. This setup lets symbolic and neural systems interact directly on shared memory without external data calls.
A key advancement lies in MORK, the scalable version of Atomspace. It supports rapid access without locking and uses efficient formats like PathMap and Merkle-DAGs to manage data. MORK also incorporates neural elements by embedding multiresolution graphs directly into memory through QuantiMORK, removing the delays that typically occur when switching between symbolic and neural processing.
We present the Hyperon AGI platform and the associated PRIMUS cognitive architecture as a unified neural-symbolic framework designed to advance from current AI to AGI to beneficial superintelligence.
Read here: https://t.co/xwxmizSnwv pic.twitter.com/O1ZuUzqtIa
— SingularityNET (@SingularityNET) December 1, 2025
PRIMUS Cognitive Loops Guiding Goal Decisions
At the decision-making level sits PRIMUS, the cognitive architecture built to run on Hyperon. It operates using two main loops. One loop manages direct goals using a mechanism called MetaMo, which controls motive generation. The other runs quietly in the background, constantly looking for patterns and improving predictions based on existing knowledge.
These loops operate on shared Atomspace memory and use common methods to control reasoning and resource use. For example, reasoning difficulty is measured by “geodesic effort,” while decision simplicity is guided by a quantale-based system.
PRIMUS includes several subsystems that perform key roles. WILLIAM compresses and generalizes patterns, while MOSES performs evolutionary program search. SubRep breaks large tasks into smaller actions. The ECAN module manages attention allocation based on utility. These systems interact fluidly, with no translation layers between symbolic and neural domains
SingularityNET Platform Driving Progress Across Fields
By 2025, Hyperon and PRIMUS are supporting game AI pilots, bioinformatics, mathematical tools, and social robotics. This approach improves not only patterns but also the framework that is being presented as a path toward broad and safe general intelligence.
The shared format makes it possible for AI models to run on android devices or across distributed networks. Code written in MeTTa can run across various systems, including blockchain-based environments. All of it is designed with reproducibility and traceability in mind, setting it apart from traditional machine learning setups.
SingularityNET’s shift from separate AI components toward a single cognitive system reflects an intent to build general intelligence with clear rules, unified memory, and decentralized execution.
Credit: Source link
