hermes-llmwiki

"Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase." — Andrej Karpathy

A Karpathy-native, zero-infrastructure memory provider for Hermes Agent. Turns your local Markdown vault into a compounding knowledge system — no Docker, no cloud, no vector DB.

Why This Exists

Every serious Hermes user hits the same wall: the agent forgets everything between sessions. Existing memory providers either:

• Lock you into a SaaS (Honcho, Mem0, Supermemory)
• Require Docker + Qdrant + Redis (Memory OS)
• Are read-only (hermes-homie-memory)
• Don't follow a structured curation model (Open Second Brain)

hermes-llmwiki is different:

Karpathy 3-Layer Architecture

┌─────────────────────────────────────────┐
│ Layer 3: Compiled Wiki (Query)          │
│  entities/  concepts/  comparisons/     │
│  projects/  queries/                    │
│         ↑  LLM curation (nightly)       │
├─────────────────────────────────────────┤
│ Layer 2: Chronicle (Daily Notes)        │
│  chronicle/daily/YYYY-MM-DD.md          │
│         ↑  auto-capture from sync_turn  │
├─────────────────────────────────────────┤
│ Layer 1: Raw (Session Exports)          │
│  raw/session-{id}.md                    │
│         ↑  on_session_end / on_pre_compress│
└─────────────────────────────────────────┘

• Layer 1 (Raw): Session dumps, temporary, disposable
• Layer 2 (Chronicle): Daily timeline, append-only, human-readable
• Layer 3 (Compiled): Atomic notes with wikilinks, cross-references, frontmatter — the "source of truth"

Install

# Option 1: pip
pip install hermes-llmwiki

# Option 2: clone to Hermes plugins directory
git clone https://github.com/chancelu/hermes-llmwiki.git ~/.hermes/plugins/hermes-llmwiki

Then activate in ~/.hermes/config.yaml:

memory:
  provider: llmwiki
  llmwiki:
    vault_path: ~/Documents/selfwiki   # or your Obsidian vault

Run setup wizard:

hermes memory setup

CLI Commands

hermes llmwiki curate              # Manually trigger: daily notes → compiled wiki
hermes llmwiki stats               # Vault statistics
hermes llmwiki search "prompt injection"   # Search compiled layer
hermes llmwiki health              # Dead links, orphans, stale notes
hermes llmwiki export --format json        # Export wiki to JSON

AI-First Note Format

Every compiled note follows Karpathy's AI-First Vault Principle:

---
type: concept
name: "Prompt Injection"
date: 2026-07-29
tags: [security, llm]
ai-first: true
confidence: high
sources: ["https://..."]
---

## For future Claude
Prompt injection is an attack where malicious input overrides the LLM's
system instructions. This page is the canonical reference.

## Key claims
- Direct injection: user input contains hidden instructions
- Indirect injection: data from external source carries payload
- Source: OWASP LLM Top 10, 2026 edition

## Related
[[LLM Security]] · [[Jailbreak]] · [[System Prompt Hardening]]

Configuration

Full config reference in CONFIG.md.

License

MIT