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

Overview

Morphological Memory is a design framework for long-horizon synthetic-agent memory, inspired by basal cognition and non-neural morphogenesis. It combines prospective goal setpoints, decay-weighted associative graphs, filesystem stigmergy and topological gating. The paper distinguishes software scaffolding from proposed native inference dynamics and recognises that contemporary agent-memory systems already support evolving records and structured retrieval.

Evidence and limitations

The companion biofield_sim suite evaluates selected mechanisms in synthetic experiments. Matched reservoir baselines are negative: a conventional Echo State Network outperforms the tested bioelectric-lattice and Physarum-flow substrates on linear memory capacity and NARMA-10.

In a 30-seed synthetic wipe-resumption world where relevance is encoded by co-occurrence, a Hebbian-decay graph attains settled recall@8 of approximately 0.61, compared with 0.25 for semantic cosine retrieval and approximately 0.05 for a structural cosine baseline. Recall drops to 0.296 shortly after a goal switch. A finite-step, goal-conditioned readout raises post-switch recall@8 to 0.754 and reduces mean simulated retrieval pages from 7.53 to 4.73. The normalisation ablation is mixed across evaluation times, and literal Physarum flux remodelling underperforms Hebbian decay.

These are bounded synthetic retrieval results, not evidence of biological equivalence, measured API savings or live-agent effectiveness. Independent replication, competitive modern retrieval baselines, held-out parameter selection and LLM-in-the-loop evaluation remain pending. The participant-observer operational note is anecdotal motivation only.

Changes in v0.7.0

• Adds the 30-seed finite-step, goal-conditioned readout experiment and its mixed normalisation ablation.
• Retains negative reservoir and literal flux-remodelling findings.
• Corrects invalid Setpoint sparsity reporting while preserving the historical export; original and reporting-corrected data are clearly distinguished.
• Revises architectural claims, related-work comparisons and bibliography, and repairs figure and PDF formatting.
• Documents local reproduction rather than independent replication. A fresh patched run matches exported recall, page-count and sparsity metrics exactly; selectivity differences are at most 1.33 × 10⁻¹². Internal state trajectory equivalence is not claimed.

Code and reproducibility audit

Companion repository, benchmarks and revision audit

Source snapshot for this revision

Commit efac04c

Qualitative browser visualiser (not live-agent telemetry)

Interactive simulation

Licensing

Paper: Creative Commons Attribution 4.0 International (CC BY 4.0).
Companion code: MIT.