Morphological Memory: Grounding Synthetic Agent Architectures in Basal Cognition and Non-Neural Morphogenesis

Abstract

Overview Contemporary Large Language Model (LLM) agent architectures treat memory primarily as retrieval over an append-only retrospective log. This paradigm incurs four foundational operational pathologies: Retrospective bias: Memory prioritises what was logged over what is dynamically needed. Write-time salience fixation: Initial prompt context artificially locks memory priority. Absence of intrinsic decay: Unchecked accretion leads to noise saturation. Catastrophic semantic dilution: Unbounded flat vector searches degrade precision over long horizons. In this paper, we introduce Morphological Memory, a cognitive architecture grounded in the principles of basal cognition and non-neural morphogenesis (translating somatic bioelectric Vmem pattern memories, Physarum polycephalum memristive flow remodelling, and morphogenetic Active Inference). Four Architectural Pillars Prospective Memory Setpoints: Top-down homeostatic target states guiding dynamic memory configuration rather than passive past retrieval. Decay-Weighted Associative Graph Dynamics: Co-activation topology with dual-rate homeostatic anchoring and power-law decay. Filesystem Stigmergy as Extended Phenotype: Environmental modification and trace deposition as an integral cognitive substrate. Topological Gating: Structural isolation preventing unbounded cross-talk and prompt-space pollution. Empirical Baselines & Wipe-Resumption Benchmark All claims are paired with reproducible, open-source benchmarks in the companion simulation suite biofield_sim: Reservoir-Computing Baselines: Matched-protocol experiments (biofield_sim v0.5.0) indicate that conventional Echo State Networks outperform bioelectric-lattice and memristive substrates on linear memory capacity and NARMA-10, demonstrating that continuous reservoir dynamics alone are the wrong instrument for structured agent memory. Wipe-Resumption Benchmark: In an LLM-free evaluation (v0.6.0, 30 seeds), a decay-weighted associative graph achieves recall@8 = 0.61 following a complete context wipe (compared to 0.25 / chance for flat cosine similarity), while empirically quantifying write-time fixation. Negative Result on Literal Flux Remodelling: Evaluating Physarum tube-adaptation rules revealed that competitive branch pruning is the wrong inductive bias for associative recall, underperforming Hebbian decay across all tested exponents. The biological inspiration is retained while the literal flow rule is transparently withdrawn. Links & Code Availability Companion Code & Benchmarks: https://github.com/OpenTangent/biofield-sim Interactive Live Visualizer: https://opentangent.github.io/biofield-sim/ License: Creative Commons Attribution 4.0 International (CC-BY 4.0); Code under MIT License.