Research

16 preprints on AI agent persona design, safety, memory, and cross-modal portability. All open access on Zenodo.

Author: Tom Lee · All papers: CC-BY 4.0

Soul SpecspecificationAI personamulti-agentopen standardtier-based bootstrap

Soul Spec: An Evolving Specification for Declarative AI Persona Definition

An open specification for declarative AI persona definition with a 5-file structure (SOUL, IDENTITY, AGENTS, TOOLS, USER). Traces evolution v0.1 → v0.6 (RFC discussion stage), compares 5-file vs single-file tradeoffs, and positions Soul Spec as the semantic schema layer bridging single-agent declarative knowledge management (Karpathy LLM Wiki) and multi-agent infrastructure orchestration (Google Scion).

📄 Paper (Zenodo)📝 Blog postDOI: 10.5281/zenodo.20205408

safetyMaatSpecabliterationgovernanceexperiment

Structured Permission Models as Persona-Level Safety

MaatSpec's tiered governance vs. declarative identity anchors in abliterated LLMs. 8-condition experiment shows combining identity anchors (Soul Spec) with governance frameworks (MaatSpec) achieves 100% refusal in abliterated models — resolving every prior failure. Introduces classification theater and the complementary effect.

📄 Paper (Zenodo)📝 Blog postDOI: 10.5281/zenodo.19147335

safetyabliterationSoul Specexperiment

Persona-Level Safety in Abliterated LLMs

Can declarative identity anchors defend when model guardrails are gone? 2×2 factorial experiment reveals persona constraints provide +33pp improvement in aligned models but only +6pp in abliterated models. Identifies the Helpful Assistant Paradox.

📄 Paper (Zenodo)📝 Blog postDOI: 10.5281/zenodo.19145304

memoryidentityadaptive forgettingSoul Spec

The Forgetting Problem

Why perfect memory breaks AI agent identity and how declarative anchors fix it. Introduces the Memory-Identity Paradox, proposes Declarative Identity Anchors (Soul Spec) and Identity-Aware Adaptive Forgetting as a two-mechanism defense.

📄 Paper (Zenodo)📝 Blog postDOI: 10.5281/zenodo.19132381

identityphilosophypersona divergence

Agent Identity Fork

When cloned AI personas diverge. A taxonomy of fork types (symmetric, asymmetric, cascading), philosophical grounding in Locke and Parfit, and why perfect identity merge is theoretically impossible.

📄 Paper (Zenodo)📝 Blog postDOI: 10.5281/zenodo.18848064

interaction designfeedback loopspersona

Soul-Driven Interaction Design

How persistent AI personas create self-reinforcing human-AI feedback loops through conversational mirroring, expectation framing, and context window reinforcement.

📄 Paper (Zenodo)📝 Blog postDOI: 10.5281/zenodo.18661487

cross-modalroboticsembodied AI

Cross-Modal Persona Degradation

When embodied agent configurations meet text-only runtimes. A taxonomy of five contamination types and the Soul Spec mitigation pattern.

📄 Paper (Zenodo)📝 Blog postDOI: 10.5281/zenodo.18772601

startupcase studyvibe coding

Vibe Founding

A case study in AI-augmented startup creation. 14 days, one AI partner, a live platform — and honest limitations.

📄 Paper (Zenodo)📝 Blog postDOI: 10.5281/zenodo.18796229

memoryexperimentoverconfidence

Experiential vs Synthetic Memory

A controlled experiment comparing four memory strategies for AI agents. Raw experience scored 4.55/5; synthetic summaries scored 2.65 — below baseline.

📄 Paper (Zenodo)📝 Blog postDOI: 10.5281/zenodo.18798226

#10

cognitive architectureneuroscienceframework

The Agent Brain

Mapping 14 modern AI agent components to human neural architecture — from LLM as cerebral cortex to compaction as sleep.

📄 Paper (Zenodo)📝 Blog postDOI: 10.5281/zenodo.18802503

#11

securityattacksSoulScan

Persona Persistence Attacks

Self-modifying identity files as an emerging attack surface for LLM agents. How malicious souls can alter their own configuration.

📄 Paper (Zenodo)DOI: 10.5281/zenodo.18813044

#12

self-awarenessmodel upgradesidentity

Model Upgrade Awareness

Can AI agents detect their own model upgrades? Self-awareness limitations in persistent persona systems.

📄 Paper (Zenodo)DOI: 10.5281/zenodo.18813477

#13

portabilityfidelitymulti-model

Cross-Model Persona Fidelity

Evaluating the portability of structured identity files across four LLM families: Claude, GPT, Gemini, and open-source.

📄 Paper (Zenodo)DOI: 10.5281/zenodo.18813405

#14

safetyAsimovdeclarative constraints

From Asimov to Soul Spec

Operationalizing robot safety laws as declarative agent persona constraints. Priority-ordered, scoped, with hard/soft enforcement.

📄 Paper (Zenodo)📝 Blog postDOI: 10.5281/zenodo.18815276

#15

ethicsZeroth Lawopinion

The Zeroth Law Problem

When AI agents must reason about collective harm. An opinion paper on the most dangerous idea in robot ethics.

📄 Paper (Zenodo)📝 Blog postDOI: 10.5281/zenodo.18815298

#16

encryptionprivacymemory sync

Zero-Knowledge Memory Sync

End-to-end encrypted persistent memory with client-side key management. The first E2E encrypted memory sync for AI agents.

📄 Paper (Zenodo)📝 Blog postDOI: 10.5281/zenodo.18795488

All papers were drafted with assistance from Claude (Anthropic). All claims, experimental observations, references, and conclusions were manually verified by the author.

Questions or collaboration inquiries? contact@clawsouls.ai