A Developer's Guide to Cognitive Architecture for Autonomous Agents
The Problem
You're building an LLM-powered agent. It sounds great for fifteen turns, then drifts into a generic assistant voice, hallucinates its own backstory, and agrees with everything the user says. You've tried longer system prompts, retrieval-augmented memory, persona reinforcement. It still breaks.
The fix isn't a better prompt. It's architecture.
The Thesis
LLMs are not cognitive architectures. They are translation layers — extraordinarily powerful ones — that convert structured internal state into natural language and back. The architecture goes around the LLM, not inside it. You build memory, motivation, emotion, planning, personality, and social cognition as deterministic, inspectable, testable code. Then you use the LLM to translate between that structured state and human conversation.
What's Inside
| Chapter | Topic |
|---|---|
| 1 | The Architecture Problem — Why three fields that don't talk to each other all have pieces of the solution |
| 2 | Perception & Attention — Biased competition, TVA resource allocation, prediction error as the fundamental signal |
| 3 | Working Memory — Why the context window is not working memory, and seven strategies for managing it |
| 4 | Memory That Forgets — Ebbinghaus curves, reconstructive distortion, mood-congruent retrieval, memory security |
| 5 | Motivation & Drives — Hull's drive theory modernized, The Sims autonomy system as case study, intrinsic curiosity |
| 6 | Emotion as Architecture — OCC appraisal + ALMA mood layers + PAD space + Sprott damped oscillators |
| 7 | Decision-Making — GOAP, behavior trees, prospect theory for irrational planning, cognitive biases as perturbations |
| 8 | Imagination & Learning — Constructive episodic simulation, prospective memory, sleep consolidation |
| 9 | Personality & Identity — Big Five as system-wide modulation, persona drift measurement, anti-drift architecture |
| 10 | Social Cognition — Bayesian Theory of Mind, CiF social exchanges, shared intentionality |
| 11 | Language & Pragmatics — Grice's maxims, Rational Speech Acts, the complete LLM translation layer |
| 12 | Temporal Cognition — Boredom as computation, patience as opportunity cost, music cognition for behavioral pacing |
| 13 | Integration & Failure — Cross-subsystem failure cascades, the failure interaction matrix, graceful degradation |
| 14 | Debugging the Synthetic Mind — Psychopathia Machinalis taxonomy, believability metrics, testing harness |
| 15 | Consciousness & Ethics — Six frameworks, the Butlin scorecard, regulatory landscape, responsible design |
Every chapter includes working Python/JavaScript implementations and ends with an honest assessment of what's proven versus speculative.
Who This Is For
Developers building autonomous agents that need to sustain identity, maintain coherent behavior over time, and feel like something more than a language model wearing a costume. Game NPCs, digital companions, simulation agents, interactive fiction characters.
If your agent lives for a single conversation and answers questions, this book is overkill. If your agent needs to remember, want, feel, decide, and remain itself across hundreds of interactions — this is what comes after prompt engineering.
Who This Is Not For
Chatbot builders. The constraints are different and the literature is different. You don't need an emotion engine for a customer service bot.
Citation
If you reference this work:
The Synthetic Mind: A Developer's Guide to Cognitive Architecture for Autonomous Agents. 2026.
Found an Error?
Open an issue with the chapter number and the specific claim. Include a source if you have one.