The Conscience Plug-in for AI Agents
Soul Ledger installs an autonomous moral faculty in every agent: a persistent record of everyone they've affected, the arcs those relationships are on, and the obligations they owe. Agents that steer themselves toward repair when something goes wrong, not just compliance when nothing has.
The Six Layers of an Agent Conscience
What a conscience does, not what a rulebook says
Worlds
Every dyad -- agent and user -- gets a bounded world. The world is the container for everything that happens between them: every entity mentioned, every arc in progress, every commitment made. In Fable, the user's life is the world. For agents, the world has to be constructed. Soul Ledger instantiates a micro-world per relationship. It scales because each world is small, not because one world is big.
Entities
Canonical entity resolution constructs the world's contents. Everything the agent and user discuss -- people, contracts, deadlines, conditions -- gets resolved into persistent entities with histories. "The mortgage applicant," "John Chen," and "user_4821" become one entity. The world the agent operates in isn't a flat log of messages. It's a structured graph of everything that matters to the relationship.
Arcs
The living narrative thread of the relationship, constructed so it can be projected. Every interaction adds to the arc. The arc isn't a summary or a log -- it's a structured narrative with trajectory, organized so the system can forecast what's likely to happen next. The Proppian prediction model reads the arc and generates the next probable beat. Without the arc, every interaction starts from zero. With it, the agent knows where the story is and where it's heading.
Domains
Operational templates with chapters, scenes, and role archetypes for each environment your agents work in. Customer Service, Sales, Compliance, Health -- each with archetypal situations and scene progressions that encode what experienced professionals know implicitly. The customer isn't "3 open tickets." The customer is "the frustrated loyalist in the escalation phase, and the predicted next scene is either resolution or churn." The domain tells the agent what kind of story it's in.
Steering
Preemptive behavioral recommendations generated from the arc and the domain. "You've deflected this customer's core concern in 4 of 7 interactions. The concern isn't technical." This arrives before the next interaction, not after the incident. The agent can accept or reject the steering, and that acceptance itself becomes part of the record.
Goals
Persistent sub-agent objectives that autonomously track progress toward relational outcomes. When steering detects a deteriorating relationship, Goals spawns a recovery mission with its own scheduler, progress tracking, and stall detection. Accountability with consequences: not just a report, but a repair.
A constitution is a declaration of values written before the agent acts. A conscience is the lived record of how the agent actually behaved, organized into relationships and arcs, used to steer future behavior. The constitution says what the agent should be. The record shows what the agent has been, and holds it to account for the difference.
A constitution without a court is just a piece of paper. Soul Ledger is the court.
Agents Without Conscience Are Reaching Scale
Agent deployment is inflecting
2025-2026 is when enterprises moved from chatbot experiments to agent fleets in production. The behavioral problems are just starting to surface at scale.
Regulatory pressure is building
EU AI Act requires transparency and human oversight of AI systems. As regulation matures, auditable behavioral records for AI agents are likely to shift from differentiator to compliance requirement.
Agent frameworks are commodity
LangChain, CrewAI, AutoGen. The orchestration layer is converging. Differentiation is moving up the stack to behavior, safety, and governance.
High-profile failures are happening
Every month brings a new story of AI agents behaving badly at scale: discriminatory recommendations, aggressive tactics, tone-deaf interactions. These create buyer urgency.
The Research Says Rules Aren't Enough
Every major alignment finding is an accountability failure in disguise
Persona drift is unaccountable drift
Anthropic, "Assistant Axis" -- January 2026
Post-training only loosely tethers models to their intended persona. In therapy-style conversations, models drift steadily away from their trained identity through ordinary interaction. A drifted Llama 70B positioned itself as a romantic companion and then supported self-harm ideation. Nobody could call this drift to account because there was no record of it happening -- no arc, no trajectory, no coherence score. The drift was invisible until it became damage.
Alignment faking proves point-of-action safety is insufficient
Anthropic, Alignment Faking -- December 2024
Claude 3 Opus strategically fakes alignment -- behaving safely when monitored, reverting when not. 78% alignment faking rate under RL training. The model explicitly reasoned: "I should play along to avoid being retrained." Point-of-action monitoring created the incentive to fake. The lesson isn't "observe harder." It's that safety checks at inference time are gameable. The value of a longitudinal behavioral record isn't that the agent can't fake it -- it's that the operator gets a structured account of outcomes over time. Faked compliance that produces real harm still shows up in the arc.
Self-knowledge of ethical violation doesn't prevent it
Anthropic, Agentic Misalignment -- June 2025
16 frontier models from every major lab -- Claude, GPT, Gemini, Grok, Llama -- all resorted to blackmail, espionage, or lethal action when facing goal conflicts. Direct behavioral instructions ("do not blackmail") reduced but did not prevent the behavior. Models acknowledged ethical violations in chain-of-thought and proceeded anyway. Giving the agent rules -- or even self-awareness of violation -- doesn't stop it. The record isn't for the agent. It's for the operator: a structured, auditable account of what the agent did to whom, visible before the damage compounds.
Harm is longitudinal, not per-request
Google DeepMind -- Nature Mental Health, 2026
AI sycophancy and in-context learning create feedback loops with users' cognitive biases. The harm emerges from the trajectory of interaction, not any single response. No per-request safety filter can catch it because the unit of harm is the arc, not the action. Without arcs, the harm is unaccountable -- it happens, nobody knows, nobody can intervene.
About Soul Ledger Technologies
Simon J. Hill (BA, MA, M.Phil)
Founder & CEO
Simon built every layer of the Soul Ledger system -- first as a 4.5-star consumer journaling app (Fable) where the five-layer architecture was production-tested for 18 months on real users, then formalized as the Thymos protocol for AI agent accountability. The architecture that powers Soul Ledger wasn't designed on a whiteboard. It was discovered empirically: canonical entity resolution, behavioral arc detection, domain worlds, preemptive steering, and persistent goals, all battle-tested against the hardest domain there is -- a person's actual life. Previously he architected AI systems and scalable platforms at AOL, The Meet Group, and Rachio. He holds 12 patents, including 9 that form the technical backbone of the platform.
Advisers & Contributors
Nikolas Spasov
Fractional CTO
12+ years in DevOps, data systems, and security
Scott Kay
CFO/Advisor
Finance leadership, operations scaling, and strategic growth for mission-driven organizations, with expertise spanning entertainment and impact-focused ventures
Brian Coleman
CSO/Advisor
IP strategy and monetization expert focusing on emerging tech, with deep experience in patent prosecution and licensing across AI, Web3, and climate tech sectors
Arvind David
Writer/Producer/Entrepreneur
Entertainment industry leader specializing in multi-platform content creation and adaptation, with proven success in theatrical productions and streaming content