Persistent personal knowledge bases powered by LLMs — beyond RAG.
Traditional retrieval-augmented generation (RAG) forces models to rediscover knowledge on every query.
LLM Wiki transforms the LLM into a dedicated knowledge engineer that builds and maintains a clean, interlinked, evolving Markdown second brain from raw documents.
Goal: Accumulate refined, hierarchical knowledge over time instead of starting from scratch on each interaction.
- Drop your documents, notes, or PDFs into the
raw/folder - Copy the content of
llm-wiki.md - Paste into any frontier LLM (Grok, Claude, GPT, etc.) and prompt:
"You are my LLM Wiki knowledge engineer. Build and maintain the wiki from the documents in raw/."
This implementation is directly grounded in the latest arXiv literature on AGI-adjacent memory systems:
- HiMem: Hierarchical Long-Term Memory for LLM Long-Horizon Agents (Zhang et al., arXiv:2601.06377, 2026)
- H-MEM: Hierarchical Memory for High-Efficiency Long-Term Reasoning in LLM Agents (Sun et al., arXiv:2507.22925, 2025)
- A-MEM: Agentic Memory for LLM Agents (Xu et al., arXiv:2502.12110, 2025)
- G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (Zhang et al., arXiv:2506.07398, 2025)
- ARC Prize 2025 Technical Report (Chollet et al., arXiv:2601.10904) and related AGI abstraction surveys
Inspired by @karpathy’s original LLM Wiki concept.
Follow on X | Open an issue
Contributing: Add real documents to raw/, run the prompt, and open a PR with improvements to the wiki/ folder.