AI · Framework
LLM Living Knowledge Base Pipeline
Architecture framework for building an LLM-maintained markdown wiki that compounds knowledge
A conceptual framework based on Andrej Karpathy's method for organizing research into a persistent, self-updating markdown knowledge base. Instead of relying on RAG or infinite context windows, an LLM compiles raw sources into structured markdown files, queries them, and files its own outputs back into the wiki. The user steers and inspects the knowledge base via an IDE like Obsidian while the model writes and maintains the index.
The framework
Main pipeline: (1) Sources — articles + papers, repos + datasets, images + diagrams; (2) raw/ — unprocessed files stored as-is, local images; (3) WIKI — .md knowledge base: summaries + backlinks, concept categories, auto-maintained index (~100 articles / ~400K words); (4) Q&A Agent — complex questions against full wiki, no RAG needed; (5) Output — markdown files, Marp slide decks, matplotlib plots. Feedback loop: outputs are filed back into the wiki — knowledge compounds. Support layer (all connected to the wiki): Obsidian (IDE frontend for wiki, view raw + compiled data, Marp plugin for slides; human reads, LLM writes), Lint + Heal (find inconsistent data, impute missing info, suggest new articles, web search to fill gaps), CLI Tools (search engine over wiki, web UI for direct use, CLI interface for LLM; vibe-coded, keeps growing). Looking ahead: finetune LLM on wiki data — knowledge in weights, not just context. Core insight: 'You never write the wiki. The LLM writes everything. You just steer — every answer compounds.' Tweeter's additions: agents that own their knowledge layer don't need infinite context windows — just good file organisation and reading their own indexes; cheaper, more scalable, more inspectable than stuffing everything into one prompt.
What it's for
- Design a persistent memory layer for autonomous AI agents
- Organize large collections of research papers, repositories, and raw notes into a compounding knowledge base
- Structure an Obsidian vault that an LLM automatically maintains, links, and lints
- Prepare structured domain datasets to fine-tune future local models
Key points
- Four-stage pipeline: raw ingestion, LLM-compiled markdown wiki, direct Q&A agent, and multi-format outputs
- Feedback loop where new answers, slides, and plots are indexed back into the wiki
- Obsidian frontend integration acting as the human inspection layer
- Automated 'lint and heal' maintenance to resolve inconsistencies and fill gaps using web search
- CLI and search tooling built specifically for model and user retrieval
Pricing
Free
Details
- Type
- Framework
- Category
- AI
- Topic
- Knowledge Management
- Open source
- No
- Platforms
- other
- For
- AI developers, researchers, and knowledge workers designing agentic workflows or personal knowledge systems
- Alternative to
- Vector RAG pipelines, MemGPT
- Tags
- knowledge base, llm agents, obsidian, markdown, architecture, second brain
Why it was saved
- @jumperz: A complete reference architecture for an AI second brain: ingest, compile, query, file back — usable as a blueprint for agent memory systems
Related
Connected workspace for notes, docs, databases, and AI automation.
A workflow for using LLMs to compile and maintain personal wikis
Ad-free platform for saving, organizing, and connecting ideas into channels
Local-first IDE and framework for building typed, agentic knowledge graphs
A design pattern for building LLM-maintained personal knowledge bases in markdown
Turn Twitter bookmarks into a searchable AI knowledge base with citations
Updated