Mathematics of Pseudo-Cognitive Dynamics: Persistent Memory, Error Learning, and Verifiable Mathematical Research in AI Systems
Year: 2026
Type: preprint
Abstract: Mathematical research with language models requires continuity across finite contexts, preservation of failed arguments, reliable retrieval of assumptions, and evidence appropriate to each proof claim. We introduce Mathematics of Pseudo-Cognitive Dynamics as a proposed framework for studying the interaction between an active model context, called the conscious layer, and persistent external memory and control mechanisms, called the subconscious layer. These terms describe information flow and control; they do not assert subjective experience. Math Brain provides a concrete case study comprising typed mathematical objects, signed association graphs, mistake memories, retrieval hooks, consolidation proposals, audit records, and interfaces to Lean. An artifact-based study of the available repositories finds substantial value for recalling project-specific knowledge and preserving corrections, together with important limits on claims of general reasoning improvement. A recent two-agent comparison scored 27/27 versus 1/27 on retained project-result questions and 27/27 in both arms on safeguard questions, at reported list-price costs of $22.7979 and $4.3693. A separately added four-arm operational study favors direct instruction delivery, while its retrieval follow-up lacks downstream agent outcomes. Other proof-audit comparisons show small gains, null results, or contamination; an earlier favorable result was undermined by a placebo control. The archived usage ledger contains 375,282,198 output tokens and 123,820,924,963 input-token events over 32 days, mostly cache reads, but does not identify the all-time construction cost of Math Brain. Research case studies distinguish complete candidate arguments, partial results, rejected routes, and problems remaining open. Formal verification is scoped to the exact checked statement and cannot substitute for statement fidelity. We develop conditional mathematical results for bounded retrieval, stable activation, correction propagation, selective warnings, and cost amortization, and specify an evaluation protocol separating knowledge access, inference, reliability, and resource cost.
Author: Shmalo, Yitzchak
Institution: Hebrew University of Jerusalem
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225910305
