Research and Papers
Peer-reviewed scientific literature, algorithmic formalizations, and empirical benchmark evaluations of compilation-based knowledge loops.

A research paper introducing a memory-centric agentic system for the complete scientific research lifecycle. Structures long-term knowledge and project artifacts into schema-governed Markdown vaults (SciMem), utilizing DAG-based multi-agent execution and continuous feedback loops (SciEvolve) to iteratively update and patch scientific knowledge over time.
Beyond Memory: A Templated Substrate for Heterogeneous Collaborative Knowledge Work with LLM Agentsarxiv.orgA research paper formalizing a templated LLM Wiki substrate for collaborative knowledge work across multi-human, multi-agent, and multi-domain environments. Details how an append-only convention preserves dead ends, negative results, and walked-back claims, introducing empirical case studies on retroactive evidence audits, failure-path preservation, and agent honesty.
Beyond RAG: LLM Wikis as Living Semantic Memory: Patterns, Empirical Findings, and Open Problemszenodo.orgA position paper analyzing the LLM Wiki pattern as compiled semantic memory, detailing scale limits and symbolic logic integration.

A preprint proposing an agent-native memory architecture that organizes knowledge into a hierarchical Context Tree (Domain, Topic, Subtopic, Entry) on the local filesystem. Integrates an Adaptive Knowledge Lifecycle (importance scoring, recency decay) and 5-tier progressive retrieval, achieving state-of-the-art results on LoCoMo without external vector databases or embedding services.
Codebase-Memory: Tree-Sitter-Based Knowledge Graphs for LLM Code Exploration via MCP (Vogel et al.)arxiv.orgA research paper formalizing the design and evaluation of Codebase-Memory. Explores constructing tree-sitter based AST call and dependency graphs exposed via MCP, demonstrating a 10x token reduction and 2.1x fewer tool calls compared to standard codebase explorers.
Compiled Memory vs Governed State: Karpathy's LLM Wiki and the Reflexive Laboratory (Peter Bell)zenodo.orgA research paper analyzing the division of cognitive labor and transaction costs in AI-assisted research workflows. Contrasts static compiled memory (LLM Wikis) with governed state environments (the Reflexive Laboratory), introducing a two-thread strategy of Master Strategy and Transcript Governance.

A foundational paper addressing compounding structural decay in long-lived agent knowledge bases. Proposes a reinforcement learning framework that interacts with a pre-constructed wiki to perform abductive error diagnosis and localized refinement actions, optimized via a Gain-Beyond-Draft (GBD) reward policy to heal broken links, coreferences, and factual anomalies.

An empirical pilot study evaluating the LLM Wiki pattern against RAG in exception-dense wastewater diagnosis. Finds that no wiki variant outperformed strong RAG, and characterizes three maintainer degenerations: fragmentation, atomization, and over-merging.

A scholarly paper framing LLM Wikis through library and information science (LIS) theory as semi-persistent knowledge organization layers. Contrasts answer-centered RAG with wiki-mediated search across four core functions (representational scaffolding, exploratory navigation, provenance exposure, and human-AI knowledge maintenance) to support accountable, inspectable scholarly discovery.

An empirical econometrics study applying the Agentic ROI framework to evaluate the compounding token economics of LLM Wikis on Qing Claw (a C# multi-agent runtime). Treats tokens as capital assets rather than consumables, detailing how a persistent markdown layer achieves 84.6% cumulative query token savings over classic RAG.

A theoretical position paper framing the architectural shift from traditional data engineering to "knowledge architecture" for AI agents. Outlines a taxonomy mapping classical guarantees (ETL, CDC, materialized views, lineage) to agent-ready patterns (knowledge ingestion, change detection, knowledge views, provenance), positioning LLM Wikis and OKF as foundational examples.

A clinical research study utilizing a persistently maintained LLM wiki as a "Living Constraint Oracle" (LCO) to govern safe offline reinforcement learning policies. Demonstrates how converting narrative clinical guidelines into weighted constraints allows IQL policies to dynamically update and prevent hazardous clinical behavior (e.g., insulin over-administration) without retraining.
LLM-Compiled Wikis for Cross-Institutional Policy Retrieval: A Benchmark on Chinese University Grant Policiesfaith.futuretechsci.orgAn empirical study introducing the T3a benchmark to evaluate LLM-compiled wikis on tabular, cross-institutional policy comparisons, demonstrating superior data consistency and completeness (the 4C framework) over standard vector RAG.
Managed Memory Is Still Not Governed State: Production Lessons from LLM Wiki v2 (Peter Bell)zenodo.orgA research paper analyzing production extensions to Andrej Karpathy's LLM Wiki concept (confidence scoring, supersession, hybrid search, audit trails, and collaboration hooks). Argues that while these extensions create a robust managed memory system, they still fall short of a governed, state-bearing research system with explicit authority and validation representation.

An empirical research paper presenting a controlled head-to-head comparison between vector RAG and a Karpathy-style LLM-compiled wiki across an authentic machine learning course corpus. Demonstrates that while both architectures achieve comparable single-fact recall, the compiled wiki dramatically outperforms RAG on cross-unit concept linking and source grounding (9.93 vs. 8.14 score; 100% vs. 64% grounding) with full citation provenance back to source lectures.

A preregistered ablation study testing progressive disclosure patterns on a 709-page markdown wiki. Demonstrates that structuring wikis with compact indexes and one-line summaries yields a 33% to 50%+ query cost reduction with non-inferior answer quality, as agents take fewer tool turns and fetch more targeted pages.

The foundational paper formalizing the Karpathy LLM Wiki design pattern into a production architecture. Introduces the Three Principles of Agent-Native Retrieval (Compilability, Composability, Evolvability) and details a system that converts unstructured documents into interlinked markdown pages, utilizing a persistent "Error Book" loop for cross-batch structural self-correction and multi-hop tool traversal.

A preregistered empirical benchmark evaluating a chunk-vector RAG baseline against an LLM-compiled Markdown Wiki. Demonstrates that while the Wiki architecture exhibits an extreme query-time token cost asymmetry, it provides vastly superior cross-paper synthesis and unprecedented claim-level citation accuracy compared to single-round similarity lookup frameworks.
Streaming Knowledge Compilation: Proactive Materiality-Scored Pinning for Time-Evolving LLM Wikisarxiv.orgA research paper formalizing Streaming Knowledge Compilation (SKC) to manage continuous document streams inside a static token cache budget. Introduces a materiality scoring algorithm ($\phi_t$) to proactively cache relevant updates, validating it across financial volatility and Wikipedia edit streams to resolve pre-training evaluation biases.
The Components of Agent Memory (Anthropic)anthropic.comA research paper analyzing long-term agent memory architectures and offline consolidation ('dreaming'). Details how models structure persistent markdown notes and Karpathy LLM wikis across sessions, benchmarks multi-session learning loops on CL-Bench 1.0 and Pokémon Red, and evaluates 128-session wiki compilation over a 350,000-document corpus.

A preprint analyzing scaling failures in the LLM Wiki pattern once vaults exceed 1,000 files, caused by index bloat and self-referential loops. Proposes the "Provenance-First" model where quotes are pointers to immutable source documents, and defines five invariants and a falsification protocol to enforce data integrity.

A research paper proposing an evolution trajectory memory framework for dialogue agents. Introduces an intermediate Memory Wiki layer that incrementally compiles conversational history into structured, interlinked wiki pages, utilizing explicit claim mutation operations (ADD, REVISE, DEPRECATE) and hierarchical routing to ground answers on LoCoMo and MedMT benchmarks.

A research paper proposing a foundation model architecture tailored for scalable, agent-native LLM Wiki representation and retrieval. Formalizes a Wiki Graph schema bridging fine-grained topological linkages with dense markdown contexts, introduces query-conditioned attentive message passing, and implements a zero-copy NCCL boundary exchange protocol achieving a 10.5x distributed training acceleration across agent long-term memory benchmarks.
WiCER: Wiki-memory Compile, Evaluate, Refine Iterative Knowledge Compilation for LLM Wiki Systemsarxiv.orgA foundational paper characterizing the "compilation gap" and attention dilution in long-context KV cache LLM Wiki deployments. Proposes an iterative, CEGAR-inspired compilation loop that automatically evaluates workspaces against diagnostic probes, isolates dropped facts, and injects them as cumulative structural preservation constraints, reducing catastrophic failure rates by 55%.

A research paper introducing a framework that co-evolves agent skills with a persistent knowledge wiki. Separates raw execution experience, accumulated wiki knowledge, and executable skills, demonstrating that persistent wiki compilation enables scalable, transferable skill evolution across diverse models and benchmarks.
