FactKaji: An Incremental Knowledge Compiler for AI Agents

Note

This analysis is based solely on the self-reported, unverified description provided by the author. No external corroboration exists for any claims made in this document.

The description states that FactKaji is an incremental knowledge compiler designed to transform evolving Markdown documents into trusted, traceable knowledge for AI agents.

It operates through a series of steps:

• Artifact ingestion (Markdown sources like ADRs, design notes, etc.)
• Event extraction (structured or hybrid model-based)
• State reduction (canonical subject state determination with conflict detection)
• Knowledge View compilation (task-specific bounded context)
• Incremental compilation (recompiles only affected artifacts)

It also includes:

• CLI and MCP server interfaces
• Dashboard for dispute inspection and settlement
• Deterministic rebuilds using content-addressed caches

Claim

The product is described as a compiler, not a search engine.

Inference It aims to improve AI agent reliability by managing evolving knowledge more like software than documents.

The author positions FactKaji as an alternative to traditional Retrieval-Augmented Generation (RAG), which they argue fails when dealing with dynamic knowledge.

Key claims:

• Knowledge should be compiled like software instead of retrieved like documents.
• It tracks how knowledge changes over time and distinguishes between decisions, corrections, rejections, etc.
• It provides a smaller and more reliable context for AI agents.
• The system supports deterministic builds even when LLMs are used during extraction.

Claim

"What if knowledge were compiled like software instead of retrieved like documents?"

Inference This reflects a shift from document-based retrieval to state-aware compilation, suggesting a new paradigm in agent knowledge management.

The description does not explicitly name target customers or personas. However, it implies usage by:

• AI agents (coding, research, enterprise assistants)
• Developers working with evolving documentation
• Teams managing large volumes of technical knowledge

It also mentions support for:

• Obsidian-style vaults
• Multi-note knowledge vaults
• Git-native workflows

Claim

The system targets users who need reliable, up-to-date knowledge for AI agents.

Inference Likely aimed at developers and teams using AI tools in software development or research contexts.

There is no evidence of pricing, business model, or monetization strategy in the description. The project appears to be a prototype submitted to a hackathon.

Claim

Not evidenced.

Inference No indication of revenue streams, customer acquisition plans, or commercial viability.

The system is built with:

• Next.js, React, Node.js, Python, TypeScript
• Tools: ai-agents, compiler, git, incremental-computing, json, knowledge-graph, llm, markdown, rag, vector-search, mcp

Core technical features include:

• Incremental compilation that guarantees full rebuild equivalence
• Hybrid deterministic and LLM-based extraction
• Content-addressed cache for reproducibility
• MCP tools integration
• Dashboard for dispute settlement
• Support for multiple languages (including Japanese)

Claim

The system supports both deterministic-only and hybrid operations.

Inference Suggests robustness in handling varied input types and ensuring consistency.

The author reports:

• End-to-end working system with real-world dogfooding
• Evaluation using realistic-prose benchmarks
• Adversarial review by Codex and Claude
• Real-vault testing and regression suite expansion
• Multiple validation methods including retrieval, knowledge-evolution, and adversarial testing

However, there is no evidence of:

• Customers or paying users
• Revenue or ARR
• Product-market fit metrics
• Market traction beyond internal use

Claim

The system has been validated through multiple testing approaches.

Inference Indicates maturity in development but not commercial traction.

The description does not mention competitors directly. However, it contrasts its approach with:

• Traditional RAG systems
• Document search engines
• Conventional knowledge management tools

It references the broader trend of "KnowledgeOps" as an emerging discipline closer to compilation and continuous integration than conventional search.

Claim

Competes with traditional RAG and document-based retrieval.

Inference Suggests a niche in AI agent knowledge infrastructure, potentially overlapping with tools for managing evolving documentation or knowledge graphs.

• Lack of commercial evidence: No revenue, customers, or traction data provided.
• Unproven market demand: The project is described as a hackathon submission; no indication of real-world adoption.
• High technical complexity: Relies on hybrid extraction and deterministic rebuilds — may be difficult to scale or maintain.
• Limited scalability assumptions: No mention of performance at scale or multi-user environments.
• Dependency on LLMs: While reproducible, the system still uses LLMs in some modes, which introduces potential instability.

Claim

The project is not yet proven in production use cases.

Inference Risk of overstatement in claims due to lack of external validation or real-world usage.

1. What specific AI agents or use cases are you targeting with FactKaji?
2. How do you plan to scale incremental compilation for large knowledge bases?
3. Have you identified any early adopters or pilot users beyond your own testing?
4. What is the roadmap for monetization or commercial deployment?
5. Can you provide examples of actual disputes detected and resolved by FactKaji?
6. How does FactKaji handle integration with existing Git workflows or DevOps pipelines?
7. Are there plans to support other formats beyond Markdown (e.g., PDFs, databases)?
8. What are the performance implications of using LLMs in hybrid mode for large-scale knowledge compilation?