Acknowledgements

The-brain stands on the shoulders of brilliant research and inspiring open-source projects.

Pi's package catalog directly inspired the-brain's plugin ecosystem design. Their approach to filtering packages by type (extension, skill, theme, prompt), sorting by downloads, and providing one-line install commands set the standard for how AI agent ecosystems should work.

Karpathy's LLM Wiki

Andrej Karpathy's concept of an interlinked, markdown-based knowledge base for LLM research inspired the-brain's Auto-Wiki plugin. The idea that knowledge should be structured, linked, and human-readable — not buried in vector databases — shaped how we think about memory consolidation.

Research Foundations

The-brain's 3-layer cognitive architecture draws from several research areas:

Self-Predictive Memory (SPM)

The Selection Layer's Surprise-Gated Prediction Error filter is grounded in predictive coding theory and self-supervised learning research:

Predictive Coding — Rao & Ballard (1999). Hierarchical predictive coding in the visual cortex. The idea that the brain constantly predicts sensory input and only processes prediction errors.
World Models — Ha & Schmidhuber (2018). Learning compressed spatial and temporal representations to predict future states. The "surprise" signal in world models directly inspired SPM's scoring mechanism.
Bootstrap Your Own Latent (BYOL) — Grill et al. (2020). Learning representations by predicting one's own latent states. The concept that self-prediction creates robust, generalizable features inspired the Identity Anchor plugin.

Graph Memory

Semantic Memory — Quillian (1968). The first computational model of semantic memory, introducing the concept of spreading activation in semantic networks.
Graph Convolutional Networks — Kipf & Welling (2017). Semi-supervised classification with graph convolutional networks. The foundation of modern graph neural networks that inspired graph-based memory representations.
Neural Turing Machines — Graves et al. (2014). The concept of external memory with differentiable read/write operations, foundational to memory-augmented neural networks.

Local Training & Privacy

LoRA: Low-Rank Adaptation — Hu et al. (2021). Efficient fine-tuning via low-rank matrices. The-brain's Deep Layer uses LoRA to consolidate memories without full model retraining.
Federated Learning — McMahan et al. (2017). The principle that data should stay on-device inspired the-brain's local-first philosophy.
MLX — Apple's machine learning framework for Apple Silicon. The trainer-local-mlx plugin makes on-device training practical by leveraging Apple's unified memory architecture.

Agent Memory Systems

MemGPT — Packer et al. (2023). OS-inspired memory management for LLMs with virtual context. The concept of tiered memory (core vs. archival) parallels the-brain's layer system.
Reflexion — Shinn et al. (2023). Verbal reinforcement learning where agents reflect on past failures. The pattern of "correction detection" in the-brain's Graph Memory mirrors Reflexion's self-evaluation loop.
Generative Agents — Park et al. (2023). Simulacra of human behavior with memory streams. The concept of retrieving relevant memories for context-aware behavior.
Meta-Harness — Meng et al. (2026), Stanford. Self-improving coding agents via edit-predict-evaluate loops. The-brain's regression fingerprinting and MCP meta-harness tools enable these systems to detect benchmark regressions across evolution cycles.
AHE: Autonomous Harness Evolution — Fudan/Peking (2026). Closed-loop codebase improvement. The-brain serves as persistent cognitive layer, tracking edit→regression patterns across cycles via SPM and identity anchor.

Cognitive Architecture

ACT-R: An Integrated Theory of the Mind — Anderson et al. (2004). An integrated theory of the mind with declarative and procedural memory. The-brain's separation of declarative memory (graph, wiki) from procedural memory (LoRA adapters) mirrors ACT-R's architecture.
A Cognitive Theory of Consciousness — Baars (1988). Global Workspace Theory: consciousness as a global workspace broadcasting to specialized processors. The-brain's hook system (BEFORE_PROMPT, AFTER_RESPONSE) is a simplified implementation of this broadcasting pattern.
Bayesian Surprise Attracts Human Attention — Itti & Baldi (2009). Bayesian surprise attracts human attention. The SPM curator's surprise-gating directly operationalizes this finding.

Open Source Dependencies

The-brain is built with:

Bun — JavaScript runtime and package manager
Drizzle ORM — TypeScript SQL ORM
Fumadocs — Documentation framework
MLX — Apple machine learning framework
cac — Command and Conquer CLI framework
Biome — Linter and formatter

Community

Thank you to everyone who has contributed ideas, bug reports, and pull requests. The-brain is what it is because of the community around it.

See CONTRIBUTING.md for how to get involved.

Inspirations

pi.dev