PRIMUS Full

The central design thesis of PRIMUS — inherited from CogPrime and developed across two decades — is cognitive synergy: the hypothesis that human-level general intelligence requires multiple specialized cognitive processes that can call on each other for help when individually stuck, achieving efficiency gains that no single process could reach alone. This is not merely a claim that multiple modules are useful; it is the stronger claim that the inter-process interactions are where the decisive intelligence gains occur.

Cognitive Synergy Theory (CST)

CST, formalized in Goertzel (2009), begins from the observation that intelligence aimed at functioning in a community of embodied, communicative agents naturally requires six distinct but interacting memory types: declarative (facts and beliefs), procedural (executable skills), sensory (perceptual representations), episodic (experienced or imagined scenarios), attentional (importance and resource allocation), and intentional (goals and motivational state). In the PRIMUS instantiation these map respectively to AtomSpace knowledge, MOSES/GEO-EVO program trees, QuantiMORK sensory encodings, internal simulation, ECAN STI/LTI values, and MetaMo motive geometries.

CST's core claim: an AGI system must contain cognitive processes specialized for each knowledge type, plus methods for synergy between these processes — mechanisms by which a process stuck in one knowledge domain can appeal to processes in other domains for aid.

The Cognitive Schematic

CST organizes goal-directed cognition around the cognitive schematic:

\[\text{Context} \wedge \text{Procedure} \rightarrow \text{Goal} \;\langle p \rangle\]

Read: "If context \(C\) holds and procedure \(P\) is enacted, goal \(G\) is achieved with confidence \(p\)." All cognitive activity partitions into two meta-operations:

• Analysis: estimate \(p\) for a posited \(C \wedge P \rightarrow G\) relationship
• Synthesis: fill in one or two of \(C\), \(P\), \(G\) given the rest, to maximize \(p\)

In PRIMUS terms: PLN handles analysis; MOSES/GEO-EVO synthesizes \(P\) given fixed \(C\) and \(G\); concept formation and pattern mining supply new candidates for \(C\); MetaMo manages goal refinement for \(G\).

Key Synergy Pairs

BBM Ch 8 and the 2009 paper catalogue specific inter-process synergies that remain central to PRIMUS:

• PLN ↔ MOSES: When PLN inference gets stuck (no high-confidence next step among available inference rules), it defers to MOSES to learn predicates characterizing the choice space. Conversely, MOSES uses PLN-derived prior knowledge to initialize deme populations and to estimate candidate program fitness without expensive evaluation (BBM §8.6).
• ECAN ↔ PLN: HebbianLinks constructed by attention allocation guide inference control toward steps most associated with the current goal. PLN reciprocally helps ECAN extrapolate conclusions about what deserves attention.
• Concept Formation ↔ PLN: New concepts compress inference trails, making previously intractable deductions feasible. PLN evaluates and refines concepts proposed by pattern mining and blending.
• Simulation ↔ PLN/MOSES: Internal simulation provides evaluation environments for both inference hypotheses and candidate programs, supplying empirical evidence where analytic methods are insufficient.

The operational key is confidence-based stuck detection: a cognitive process considers itself stuck when it has no high-confidence estimates about its next step. Deferral is guided by PLN's indefinite probability truth values, which track both probability and confidence. This creates a self-organizing dynamic where resources flow toward whichever process can make the most confident progress.

Tricky Cognitive Synergy

BBM Ch 8 introduces the tricky cognitive synergy hypothesis: components designed for synergetic AGI are necessarily harder and more complex than standalone narrow-AI components, because each must have the internal flexibility to handle interactions with many other components. A consequence is that partial AGI systems may perform worse on any given benchmark than a simpler narrow system designed specifically for that task — making intermediate progress toward AGI inherently difficult to measure. For PRIMUS development, this means individual component benchmarks can be misleading; the decisive test is whether component integration produces emergent capability beyond the sum of parts.

From CogPrime Synergy to PRIMUS Cooperation

CogPrime's cognitive synergy relied on hand-designed inter-process bridges and engineering judgment about when and how to defer. The 2025 PRIMUS design formalizes these interactions through three mechanisms CogPrime lacked:

• TransWeave provides a certified composition algebra ensuring that learn→transfer ≈ transfer→learn across paradigm boundaries, replacing ad-hoc cross-process bridges
• Weakness theory gives each algorithm a native Occam's razor so simplicity preferences compose across module boundaries via functorial maps between compatible quantales
• Geodesic control replaces heuristic stuck-detection with a uniform geometric selection rule (\(f \cdot g\) product) that applies across PLN inference, MOSES search, TransWeave transfer, and self-modification

The specific synergy pairs catalogued in BBM — PLN↔MOSES, ECAN↔PLN, etc. — are thus not abandoned but subsumed: each is now a special case of the geodesic control + weakness + TransWeave framework operating over shared Atoms.