PyCognition: A Unified Dual-Process Neuro-Symbolic Architecture for Amortized Execution, Self-Healing Skill Refactoring, and Longitudinal Episodic Orchestration

PyCognition (https://github.com/MasterUaKi/pycognition) is an open-source, domain-agnostic neuro-symbolic orchestration framework grounded in Dual-Process Cognitive Theory. It unifies fast, deterministic procedural execution (System 1) with deliberative Large Language Model synthesis, structural reflection, and semantic arbitration (System 2) to eliminate recurring inference costs, structural web drift failures, and cross-source data fragmentation in long-lived autonomous agents. Architectural Layers: Layer 1 — Utility-Guided JIT Skill Compilation (DualProcessRuntime, MetaArbiter): Routes tasks via utility-theoretic arbitration and crystallizes repetitive algorithmic intents into verified local Python subroutines, achieving an 11,326.7× marginal latency speedup (0.292 ms mean System 1 execution) and a 38.0% immediate token reduction across N = 100 heterogeneous operational requests. Layer 2 — Telemetry-Driven Skill Refactoring & AST Sandboxing (pycognition.skills): Replaces record-level toxic micro-compensation with batch corruption telemetry (TelemetryEvaluator), declarative capability-scoped AST auditing (SkillCapability, ASTSafetyAuditor), and monotonic regression gating. Evaluated on WebDrift-Fault-50 (10 cycles), it reduces self-healing LLM calls by 94.0% with zero database poisoning and blocks 100% (10/10) of adversarial code injection payloads. Layer 3 — Bitemporal Episodic Memory & Multi-Horizon Planning (pycognition.temporal): Consolidates multilingual and cross-portal records via dual-tier entity resolution with persistent alias crystallization (NeuroSymbolicEntityResolver, AliasMemoryStore), tracks bitemporal lifecycle transitions (TemporalEventGraph), and schedules proactive actions across immediate (H1), tactical (H2), and strategic (H3) horizons (ProactiveHorizonPlanner). Evaluated on ChronoDrift-Entity-60 (60 episodes × 6 snapshots = 360 events), it achieves F1 = 1.000 in entity resolution and lifecycle transition recall while reducing resolution LLM calls by 81.5% (10 vs. 54 calls). Live Production Validation (Grant Harvester v2.2.0, DOI: 10.5281/zenodo.24106395): Deployed across 24 German and European academic funding portals (1,230 active records), PyCognition consolidated duplicate multilingual and cross-portal listings into 958 canonical entities (-22.1% redundancy) and achieved a 31.1× wall-clock speedup (from 674.7 s to 21.7 s) with zero recurring System 2 LLM calls in warm synchronization cycles. [DE] Kurzfassung: PyCognition ist ein quelloffenes, domänenunabhängiges neuro-symbolisches Orchestrierungs-Framework auf Basis der Zwei-Prozess-Theorie, das deterministische prozedurale Ausführung (System 1) mit deliberativer LLM-Synthese, telemetriegesteuertem AST-geprüftem Skill-Refactoring und einem bitemporalen Ereignisgraphen (System 2) vereint. Auf drei Benchmarks sowie einem produktiven Multi-Portal-Verzeichnis mit 1.230 Fördermitteldatensätzen (Grant Harvester v2.2.0) reduziert PyCognition wiederkehrende System-2-Inferenzkosten auf null und beschleunigt Synchronisationszyklen um den Faktor 31,1×.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24364612
Primary Topic
AI-based Problem Solving and Planning
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

PyCognition: A Unified Dual-Process Neuro-Symbolic Architecture for Amortized Execution, Self-Healing Skill Refactoring, and Longitudinal Episodic Orchestration

Oleh Riabtsev
Zenodo (CERN European Organization for Nuclear Research)
AI-based Problem Solving and Planning
article

PyCognition: A Unified Dual-Process Neuro-Symbolic Architecture for Amortized Execution, Self-Healing Skill Refactoring, and Longitudinal Episodic Orchestration

Oleh Riabtsev
article en

Abstract

PyCognition (https://github.com/MasterUaKi/pycognition) is an open-source, domain-agnostic neuro-symbolic orchestration framework grounded in Dual-Process Cognitive Theory. It unifies fast, deterministic procedural execution (System 1) with deliberative Large Language Model synthesis, structural reflection, and semantic arbitration (System 2) to eliminate recurring inference costs, structural web drift failures, and cross-source data fragmentation in long-lived autonomous agents. Architectural Layers: Layer 1 — Utility-Guided JIT Skill Compilation (DualProcessRuntime, MetaArbiter): Routes tasks via utility-theoretic arbitration and crystallizes repetitive algorithmic intents into verified local Python subroutines, achieving an 11,326.7× marginal latency speedup (0.292 ms mean System 1 execution) and a 38.0% immediate token reduction across N = 100 heterogeneous operational requests. Layer 2 — Telemetry-Driven Skill Refactoring & AST Sandboxing (pycognition.skills): Replaces record-level toxic micro-compensation with batch corruption telemetry (TelemetryEvaluator), declarative capability-scoped AST auditing (SkillCapability, ASTSafetyAuditor), and monotonic regression gating. Evaluated on WebDrift-Fault-50 (10 cycles), it reduces self-healing LLM calls by 94.0% with zero database poisoning and blocks 100% (10/10) of adversarial code injection payloads. Layer 3 — Bitemporal Episodic Memory & Multi-Horizon Planning (pycognition.temporal): Consolidates multilingual and cross-portal records via dual-tier entity resolution with persistent alias crystallization (NeuroSymbolicEntityResolver, AliasMemoryStore), tracks bitemporal lifecycle transitions (TemporalEventGraph), and schedules proactive actions across immediate (H1), tactical (H2), and strategic (H3) horizons (ProactiveHorizonPlanner). Evaluated on ChronoDrift-Entity-60 (60 episodes × 6 snapshots = 360 events), it achieves F1 = 1.000 in entity resolution and lifecycle transition recall while reducing resolution LLM calls by 81.5% (10 vs. 54 calls). Live Production Validation (Grant Harvester v2.2.0, DOI: 10.5281/zenodo.24106395): Deployed across 24 German and European academic funding portals (1,230 active records), PyCognition consolidated duplicate multilingual and cross-portal listings into 958 canonical entities (-22.1% redundancy) and achieved a 31.1× wall-clock speedup (from 674.7 s to 21.7 s) with zero recurring System 2 LLM calls in warm synchronization cycles. [DE] Kurzfassung: PyCognition ist ein quelloffenes, domänenunabhängiges neuro-symbolisches Orchestrierungs-Framework auf Basis der Zwei-Prozess-Theorie, das deterministische prozedurale Ausführung (System 1) mit deliberativer LLM-Synthese, telemetriegesteuertem AST-geprüftem Skill-Refactoring und einem bitemporalen Ereignisgraphen (System 2) vereint. Auf drei Benchmarks sowie einem produktiven Multi-Portal-Verzeichnis mit 1.230 Fördermitteldatensätzen (Grant Harvester v2.2.0) reduziert PyCognition wiederkehrende System-2-Inferenzkosten auf null und beschleunigt Synchronisationszyklen um den Faktor 31,1×.

Zenodo (CERN European Organization for Nuclear Research)
University of Vechta (DE)
AI-based Problem Solving and Planning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon