MeTTa-NARS Deep Dive

help edit space_dashboard
Draft — This content has not been approved for publication.
ai generated
open_in_full page
fullscreen modal
edit edit
space_dashboard advanced
human approved
open_in_full page
fullscreen modal
edit edit
space_dashboard advanced
Authors: Patrick Hammer
Contributors: Patrick Hammer, Peter Isaev

MeTTa-NARS is Patrick Hammer's MeTTa port of the Non-Axiomatic Reasoning System (NARS) — the reasoning tradition founded by Pei Wang — derived specifically from the OpenNARS-for-Applications (ONA) control architecture. It implements the NAL-1 through NAL-5 inference layers in MeTTa, carrying NARS's two-dimensional frequency/confidence truth values through inference. It is paradigm-distinct from PLN: the two address the same problem (reasoning under uncertainty) with different truth-value math, inference primitives, and control regimes.

Last verified: 2026-06-04 (at repo HEAD 945e2f8, 2024-10-14).

Core Mechanisms

  • NAL-1..NAL-5 inference layers. Source files implement the canonical NARS truth functions (deduction, induction, abduction, comparison, analogy) and conversion rules, carrying the frequency/confidence truth-value pair through inference per NAL semantics.
  • ONA-derived control. The control structure mirrors OpenNARS-for-Applications (Hammer's NARS implementation) — selection of beliefs by priority/durability, attention-buffer management, and time-bounded reasoning under the Assumption of Insufficient Knowledge and Resources (AIKR).
  • MeTTa-native types. Truth values, beliefs, and inference rules are expressed as MeTTa atoms; rules are implemented as MeTTa rewrite patterns. This lets NARS compose with other MeTTa code on the same runtime.
  • NARS-GPT interop (proof-of-concept). MeTTa-NARS includes a NARS-GPT integration proof-of-concept. Treat it as a POC, not production behavior.

Implementation Surface

RepositoryHEADDateCharacter
patham9/metta-nars945e2f82024-10-14Primary maintained line in Hammer's portfolio; implementation-backed for NAL-1..5. The README confirms an ONA-style NARS implementation in MeTTa with non-axiomatic logic and multi-step reasoning examples.

How MeTTa-NARS Relates to Its Neighbors

vs PLN

NARS truth values are frequency/confidence pairs (f, c); PLN truth values are strength/count pairs (s, n). The semantics, evidence-accumulation rules, inference operators, and control regimes all differ. Goertzel's Mattermost paper "Comparing Truth Value Formulas in PLN and NARS Under High Uncertainty of Node Probabilities" attempts a comparison framework, but its equations and uncertainty-mapping formulas were not found in the metta-nars source (verified by cross-grep at HEAD), and the paper itself carries a hallucination warning (it was GPT-o1-produced). Treat any "PLN ≈ NARS" claim from that paper with skepticism. More generally: PLN-comparable reasoning frameworks should be treated as paradigm-distinct, not as implementations or successors of PLN.

vs AIRIS

Different paradigms despite the shared author (Hammer co-authored AIRIS and authored MeTTa-NARS). AIRIS uses a curiosity / state-graph / AIRIS-confidence apparatus; MeTTa-NARS implements NAL inference. The two are not collapsible.

vs NACE

Shared author (patham9) and shared NAL truth-value semantics, but different roles: MeTTa-NARS is a faithful NAL-1..5 reasoner port; NACE is a causal-rule learning agent that uses NAL evidence values for tracking. NACE bridges to MeTTa-NARS / ONA, but they are separate projects (neither wraps the other).

vs MeTTa-Motto

MeTTa-Motto sits outside this cognitive-algorithm coverage (it is a cross-organization topic). The metta-nars proof-of-concept interop with Motto is a peripheral integration note, not a reason to fold Motto in here.

What This Card Is Not

  • Not a PLN implementation, successor, or equivalent. NARS frequency/confidence truth values are paradigm-distinct from PLN strength/count. The two coexist in the wiki; cross-linking is at the paradigm-boundary level, not the implementation level.
  • Not the same as NACE. NACE is a learning agent that uses NAL evidence; MeTTa-NARS is the reasoner. They are sibling projects in Hammer's portfolio.
  • Not a complete NAL-1..NAL-9 implementation. NAL-1 through NAL-5 are implementation-backed at HEAD; the higher layers (NAL-6 through NAL-9) are not in scope at this version.
  • Not the canonical OpenNARS. MeTTa-NARS is Hammer's MeTTa port derived from the ONA tradition. The C-side OpenNARS / OpenNARS-for-Applications repos are upstream history, not part of this MeTTa wiki coverage.

Status and Sources

Implementation status: Maintained research line. The latest commit ("Update: filter timing indicators", 2024-10-14) suggests recent control-loop refinement. The README confirms NAL non-axiomatic logic plus multi-step reasoning examples.

Primary sources:

  • Wang, P. (2013). Non-Axiomatic Logic: A Model of Intelligent Reasoning. World Scientific. The canonical NARS reference — full book ingested as a Publication + RawData card.
  • Hammer, P. (n.d.). MeTTa-NARS: README + repository at patham9/metta-nars.
  • view (title=NARS implementation in MeTTA) not supported for Publications+NARS implementation in MeTTA — technical report on the MeTTa-NARS architecture, logic, memory, control, temporal/procedural reasoning, and attention allocation.
  • view (title=Functional Equivalence with NARS) not supported for Publications+Functional Equivalence with NARS — ONA/NARS functional-equivalence and transfer-learning source.
  • view (title=Self in NARS) not supported for Publications+Self in NARS, an AGI System — NARS self-knowledge, self-control, mental operations, and emotion/feeling mechanisms.
  • view (title=A Model of Unified Perception and Cognition) not supported for Publications+A model of unified perception and cognition — NARS perception/cognition integration source.
  • view (title=Reasoning-Learning Systems Based on NARS Theory) not supported for Publications+Reasoning-learning systems based on non-axiomatic reasoning system theory — NARS/ONA as a reasoning-learning alternative to reinforcement-learning baselines.
  • view (title=NARS solving the bAbI tasks) not supported for Publications+NARS Solving the Facebook AI Research bAbI Tasks — concrete NARS question-answering task examples.
  • view (title=Neurosymbolic Hybrid Driver Collision Warning) not supported for Publications+Neurosymbolic Hybrid Driver Collision Warning — NARS as the adaptive neurosymbolic reasoning layer in a perception/action application.
  • Goertzel, B. (n.d.). Comparing Truth Value Formulas in PLN and NARS Under High Uncertainty of Node Probabilities. Mattermost paper. Treat as boundary input only — the paper carries a GPT-o1 hallucination warning; cross-validate any extracted claim against this card, the Pei Wang NARS bibliography, and the PLN cluster material before adopting.

Provenance: the verdicts on this card come from a 2026-05-07 source-code review; the extraction archive is at scripts/archive/non_clustered_haa_pilot/source5_NACE_AIDSL_MOSES_MeTTaNARS/.