Abstract
We present Pith, a cognitive architecture that provides persistent, governed memory for large language model (LLM) agents. Unlike stateless retrieval-augmented generation (RAG) systems or simple key-value memory stores, Pith implements a multi-layered governance stack that manages the full lifecycle of learned knowledge: acquisition, belief state management, temporal decay, contradiction detection, authority scoring, and cross-session coherence. We describe the architecture's core components — a five-state belief lifecycle, four-tier feedback system, knowledge area segmentation, and a novel behavioral benchmark (CogGov-Bench) for evaluating governance effectiveness. Deployed in production with 4,874 concepts across 3,043 sessions, Pith achieves a CogGov-Bench composite score of 69.0/100 with 100% stale knowledge resistance and 93.3% context integrity. We release this work as a defensive publication establishing prior art for governed cognitive memory architectures.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Estey-Ang, Andrew, "Pith: A Governed Cognitive Architecture for Persistent AI Memory", Technical Disclosure Commons, (March 30, 2026)
https://www.tdcommons.org/dpubs_series/9660