Agents Neuron

A Claude Code skill that turns your reading into a personal, compounding knowledge wiki inside Obsidian. Import anything — URLs, notes, pasted text — and Agents Neuron filters it through your identity profile, keeps only what matters, and builds structured wiki pages with full source traceability.

Supports multiple Obsidian vaults as sources (e.g. a personal vault and a work vault) — Agents Neuron scans all of them and writes everything into one unified wiki.

Sources (URLs, notes, text)  →  neuron add   →  raw sources (immutable)
                                    ↓
                             neuron ingest  →  score against identity filter
                                              below threshold → archived
                                              above threshold → wiki page
                                    ↓
                             neuron query   →  answers with citations
                             neuron lint    →  health checks
                             neuron filter  →  evolve relevance over time

Inspired by Karpathy's original write-up and Baljanak's learning filter. Built from scratch.

---

Getting started

Prerequisites: Claude Code, Bash 4+

git clone https://github.com/glonlas/neuron.git && cd neuron
make install                        # symlinks + seed ~/.agents-neuron/ config
# Edit ~/.agents-neuron/config.yaml # set vault_path
neuron bootstrap                      # run in Claude Code to init vault

Validate: skill/scripts/doctor.sh | Uninstall: make uninstall

---

Example

# In Claude Code
neuron add https://www.reddit.com/r/LocalLLaMA/comments/1s49lvh/gguf_llamacpp_vs_mlx_round_2_your_feedback_tested/
neuron ingest

Agents Neuron fetches the thread, scores it against your identity filter, and — if it clears your relevance threshold — creates a structured wiki page (e.g. Neuron/Comparisons/GGUF llama.cpp vs MLX.md) with inline citations back to the source.

---

Automation

macOS (recommended) — use the provided launchd helper:

./helpers/setup-launchd.sh           # install agents
./helpers/setup-launchd.sh --uninstall  # remove them

Installs three agents: daily neuron ingest at 08:00, weekly neuron lint and neuron filter evolve on Mondays at 09:00. Logs to ~/.agents-neuron/launchd.log.

Linux / cron — add to your crontab with crontab -e (use the full path from which claude):

0 8 * * *   /path/to/claude -p "neuron ingest"        >> ~/.agents-neuron/cron.log 2>&1
0 9 * * 1   /path/to/claude -p "neuron lint"          >> ~/.agents-neuron/cron.log 2>&1
5 9 * * 1   /path/to/claude -p "neuron filter evolve" >> ~/.agents-neuron/cron.log 2>&1

Pair with neuron scan in your morning terminal session to pull overnight note changes.

---

Quick reference

Command	What it does
neuron scan	Pull recently modified vault notes into the wiki
`neuron add <url	text>`
neuron ingest	Score pending sources, create wiki pages
neuron query <question>	Answer from wiki with citations
neuron lint	Health check
neuron filter evolve	Tune relevance weights from usage patterns

---

Documentation

Doc	Contents
Commands	Full command reference, daily usage patterns, recommended cadence
Configuration	config.yaml, identity filter setup, scoring dimensions
Architecture	File structure, scripts vs LLM split, page types, cross-platform support
Troubleshooting	Common errors and doctor.sh
Contributing	How to add skills, scripts, and submit PRs

---

Credits

• Andrej Karpathy — the original personal knowledge wiki concept
• Baljanak — identity-aware learning filter

License

MIT