Knowledge Management Landscape Research (2026-06-29)

Research conducted when designing the m3ta-brain vault. Condensed for future reference.

1. Karpathy LLM Wiki Pattern (April 2026)

Andrej Karpathy published a "pattern gist" for LLM-maintained personal knowledge bases.

Core idea: Instead of RAG (re-derive from raw docs every query), the LLM builds and maintains a persistent, compounding wiki — cross-references already there, contradictions already flagged, synthesis already done.

Three layers:

Raw sources (raw/) — original articles, papers, transcripts
LLM Wiki (wiki/) — entity pages, concepts, comparisons, synthesis (LLM writes ALL of this)
Schema (CLAUDE.md / AGENTS.md) — tells the LLM how the wiki is structured

Key quote: "Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase."

What we adopted: AGENTS.md as schema, structured directories, LLM-maintained notes, Obsidian as viewer. What we extended: Added project state, decisions, people, learnings — not just research concepts.

Source: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

2. Daniel Miessler PAI / Personal AI Infrastructure (v5.0, May 2026)

PAI ("Life Operating System") is the most comprehensive personal AI system design.

7 Architecture Components: Intelligence, Context (Memory), Personality, Tools, Security, Orchestration, Interface.

Key insight — Scaffolding > Model: "When Kai gives me a result I don't want, it's almost never because Claude is dumb. It's because my scaffolding didn't provide the right context."

Memory System v7.6 — three tiers by PURPOSE, not chronology:

MEMORY/WORK/ — active task tracking, ISAs (Ideal State Artifacts)
MEMORY/KNOWLEDGE/ — typed graph (People, Companies, Ideas, Research)
MEMORY/LEARNING/ — meta-patterns, signals, failures, corrections
MEMORY/RELATIONSHIP/ — DA-Principal relationship notes
MEMORY/OBSERVABILITY/ — every tool call, hook firing, satisfaction signal

Signal capture: Every interaction generates signals — explicit ratings, implicit sentiment, failure captures (ratings 1-3 trigger full-context saves).

Principles adopted: "Text over opaque storage" (if you can't cat it, don't store it), filesystem as context (no RAG), memory structured by purpose.

What we adopted: Purpose-based directory structure (Telos/Preferences/Projects/Decisions/People/Learning), write policy with review gates, learning patterns and corrections as first-class citizens. What we skipped: ISA/ISC verification system, Pulse dashboard, hooks, voice — too Claude-Code-specific.

Source: https://danielmiessler.com/blog/personal-ai-infrastructure, github.com/danielmiessler/PAI

3. qmd — Query Markup Documents (Tobi Lütke, 26K stars)

Fully local CLI search engine for markdown. BM25 + Vector + LLM Reranking, all on-device.

Three search modes:

qmd search — BM25 keyword (fast, no LLM)
qmd vsearch — Vector semantic (embeddings)
qmd query — Hybrid: query expansion + parallel retrieval + RRF fusion + LLM reranking

Key features: Collections (named indexed directories), Context (human descriptions attached to paths), MCP server mode, multiple output formats (--json, --md, --files).

Local models (~2GB total): embeddinggemma-300M, qwen3-reranker-0.6b, qmd-query-expansion-1.7B.

Our plan: Install as retrieval layer over m3ta-brain + tech wiki + nemoti wiki. Collections: brain, tech-wiki, nemoti-wiki. Not yet installed (requires Node.js/Bun on server).

Source: github.com/tobi/qmd, npm: @tobilu/qmd

4. Obsidian MCP Integration Landscape

Multiple MCP servers connect AI agents to Obsidian vaults:

Server	Transport	Key Feature
Local REST API plugin	HTTP (localhost:27124)	Built-in MCP at `/mcp/`, the foundation
yanxue06/obsidian-mcp	stdio/HTTP	Graph-aware: backlinks, multi-hop, safe rename
lstpsche/obsidian-mcp (Rust)	stdio/HTTP	Single binary, filesystem-native, works without Obsidian running
Vasallo94/obsidian-mcp-server	stdio	Explicit Hermes support documented
cyanheads/obsidian-mcp-server	stdio/HTTP	14 tools, surgical section editing, read/write path scoping

Key limitation for us: MCP over localhost requires Obsidian on same machine as agent. Hermes runs server-side, Obsidian runs on Sascha's desktop. So Git-sync is Phase 1, MCP is Phase 2 (optional).

5. Shared Memory Platforms

Platform	Architecture	Key Differentiator
Open Second Brain	Obsidian-native, Hermes-first	Nightly "dream" passes, preferences with confidence bands
Agentkeep	Git-backed markdown vault	Content-hash compare-and-swap, safe two-driver editing
Noosphere	PostgreSQL + Redis	REST API, scoped access, Obsidian export/import

What we adopted from these: Git as the sync/merge mechanism (Agentkeep), draft-then-review pattern for sensitive notes (Open Second Brain), plain markdown as the substrate (all three).

Design Decisions Summary

Decision	Chosen	Rejected	Rationale
Sync mechanism	Git (Gitea)	Syncthing	Versioned, auditable, conflict diffs, Gitea already exists
Link format	Standard Markdown links (`[text](relative/path.md)`)	Obsidian-only `[[wikilinks]]`	Works in both Gitea and Obsidian; current vault rules supersede the initial design note
Language	German + EN tech	English only	Match how Sascha communicates
Agent instructions	`AGENTS.md`	Agent-specific config	Multi-agent compatible (Pi, OpenCode, Claude Code, Hermes)
Memory model	Purpose-based dirs	Chronological	Inspired by PAI v7.6 (WORK/KNOWLEDGE/LEARNING pattern)
Retrieval	qmd (planned)	RAG/black-box	Local, hybrid search, no cloud dependency

5.7 KiB Raw Blame History