Param Siddharth paramsiddharth

LLM Wiki v2

A pattern for building personal knowledge bases using LLMs. Extended with lessons from building agentmemory 20K+ Stars ⭐️, a persistent memory engine for AI coding agents.

This builds on Andrej Karpathy's original LLM Wiki idea file. Everything in the original still applies. This document adds what we learned running the pattern in production: what breaks at scale, what's missing, and what separates a wiki that stays useful from one that rots.

What the original gets right

The core insight is correct: stop re-deriving, start compiling. RAG retrieves and forgets. A wiki accumulates and compounds. The three-layer architecture (raw sources, wiki, schema) works. The operations (ingest, query, lint) cover the basics. If you haven't read the original, start there.

LLM Wiki

A pattern for building personal knowledge bases using LLMs.

This is an idea file, it is designed to be copy pasted to your own LLM Agent (e.g. OpenAI Codex, Claude Code, OpenCode / Pi, or etc.). Its goal is to communicate the high level idea, but your agent will build out the specifics in collaboration with you.

The core idea

Most people's experience with LLMs and documents looks like RAG: you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works, but the LLM is rediscovering knowledge from scratch on every question. There's no accumulation. Ask a subtle question that requires synthesizing five documents, and the LLM has to find and piece together the relevant fragments every time. Nothing is built up. NotebookLM, ChatGPT file uploads, and most RAG systems work this way.

Working Clawdbot/Moltbot setup with local Ollama model

[Update 2026-02-02: nemotron-3-nano also performs well on same setup; see comment below]

This is a guide to setting up Clawdbot/Moltbot with a local Ollama model that actually works -- meaning it has good tool use and decent speed. The main requirement is 48GB of VRAM. I have yet to find a model that fits on less than this and still works on Moltbot.

The setup involves creating a tool-tuned variant of qwen2.5:72b and modifying a range of configs in Moltbot. At the end you'll get a local Moltbot instance that can use tools (exec, read, write, web search), read skills, and perform agentic tasks without any cloud API dependencies. On my system I get ~16 t/s and have yet to come across a tool/skill that my bot can't use.

Claude Opus wrote the first draft of this Gist, then I (a human) checked and edited it.

Why I Left AT Protocol

I'm Tree. I think and write about decentralization - analyzing power dynamics, architectural principles, and how systems can preserve or undermine individual autonomy. I've been involved with AT Protocol development and community discussions for some time.

TL;DR: I deleted my atproto account not because of bullying, though that happened. I left because the community's response to my architectural critique demonstrated exactly the power dynamic I was warning about, while simultaneously making it impossible to discuss.

The Technical Argument

Infrastructure should be separate from ideology. This is a foundational principle for resilient decentralized systems.

Cloudflare WARP tunnel via Wireguard client

This example provides a simple configuration for a Debian client to have a Cloudflare tunnel while not installing the official Cloudflare WARP client. Note: Tunnel transport outbound to engage.cloudlflare.com on udp/2408 is default, with a dynamic listening udp port and a fwmark for packet matching by wireguard. Any applicable firewall rules may need to be adjusted.

Top-level GitHub project to convert cloudflare endpoint to generic wireguard configuration file: https://github.com/ViRb3/wgcf

Install dependancies

sudo apt install openresolv wireguard-tools golang git

Get the latest client from Github and build using go

git clone https://github.com/ViRb3/wgcf.git

Install .NET 8 on Raspberry pi

Install .NET via the following script:

https://learn.microsoft.com/en-us/dotnet/core/tools/dotnet-install-script

1. Perform dry run and review envvar `DOTNET_INSTALL_DIR`

I advise to first do a dry run and to enable verbose mode to review if envvar DOTNET_INSTALL_DIR is set correctly:

	"""
	The most atomic way to train and run inference for a GPT in pure, dependency-free Python.
	This file is the complete algorithm.
	Everything else is just efficiency.

	@karpathy
	"""

	import os # os.path.exists
	import math # math.log, math.exp

	/*
	I was looking to try a non-kindle e-reader, but my library kept weighing me down. While I had calibre and the de-drm plugin, I found the downloaded books on my kindle were in kfx format and weren't playing well with calibre, even with the kfx input plugin.
	Further, I found downloading books directly from web (Amazon > content library) did give me an azw3 file, which works well with de-drm.

	To automate part of the process, I wrote a short js script. It doesn't walk through all the pages in content library, but it does download all the books currently visible on the list. Amazon lets us view 25 books per page, so I just had to switch pages, run script, repeat.
	The meat of this script is constructing the url used to start a download. Device type, serial number, customer id etc. will be different for everyone, so it's best to download one book manually and check the network inspector for the request details. Apart from the book's `key`, all other details remain same for all the books (at least they did for me

	import javax.crypto.Cipher;
	import javax.crypto.CipherInputStream;
	import javax.crypto.NoSuchPaddingException;
	import javax.crypto.SecretKey;
	import javax.crypto.spec.IvParameterSpec;
	import javax.crypto.spec.SecretKeySpec;
	import java.io.ByteArrayInputStream;
	import java.io.IOException;
	import java.io.InputStream;
	import java.nio.file.Files;

	COMPOSE_PROJECT_NAME=remote-signer
	ENABLE_API_KEY_SUPPORT=true
	APPROVED_API_KEY=( UUID4 or anything really as long as this is exported on the sender side )