Skip to content

Lineage and Credits

Memex stands on a clear lineage. The first and most important credit goes to Andrej Karpathy's LLM Knowledge Bases / LLM Wiki pattern.

Primary credit: Andrej Karpathy

Karpathy's LLM Knowledge Bases framing defines the core move this project builds on: use LLMs to maintain a personal knowledge base as a persistent artifact, not as a disposable chat transcript or one-shot RAG answer.

His LLM Wiki gist gives the clearest operating pattern:

  • raw sources are preserved as immutable evidence
  • the LLM maintains a structured markdown wiki between the user and the raw sources
  • new sources update existing entity, topic, and synthesis pages
  • contradictions, cross-references, indexes, and logs are maintained over time
  • answers can become new wiki pages, so inquiry compounds inside the knowledge base

Memex adopts this core thesis directly: the wiki is a compiled layer of knowledge, and the LLM is the maintainer.

References:

  • Andrej Karpathy, LLM Knowledge Bases, X thread: https://x.com/karpathy/status/2039805659525644595
  • Andrej Karpathy, LLM Wiki, GitHub gist: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

Operational influence: Garry Tan

Garry Tan's GStack and GBrain informed the operational shape around the Karpathy wiki pattern.

Thank you to Garry Tan for developing and sharing both GStack and GBrain. Those projects make the agent-operable version of this idea much more concrete: GStack shows how to package durable agent workflows, and GBrain shows how to turn a markdown knowledge base into a searchable, cited, graph-aware brain layer.

GStack demonstrates a markdown-native agent operating system: explicit roles, repeatable skills, review workflows, QA, security checks, release routines, and a disciplined way to turn an AI coding agent into a structured software team.

GBrain demonstrates the retrieval and brain layer: a local-first agent brain with synthesis, citations, graph traversal, gap analysis, PGLite storage, embeddings, and autonomous maintenance workflows.

Memex borrows from these projects at the architecture level:

  • skills as durable workflow contracts
  • retrieval that returns cited synthesis rather than only ranked chunks
  • graph-aware memory rather than flat note search
  • health checks, evals, and benchmarks as operating requirements
  • local-first files plus machine indexes as complementary layers

References:

  • Garry Tan, GStack: https://github.com/garrytan/gstack
  • Garry Tan, GBrain: https://github.com/garrytan/gbrain/tree/master

What Memex adds

Memex is not trying to hide its influences. Its job is to package the lineage into a public-ready framework that can be audited, forked, and adapted without leaking a private vault.

The additions here are:

  • a public/private repository boundary
  • typed entity and item schemas
  • reusable templates for sources, items, decisions, relationships, and reviews
  • Agent Skills-compatible workflows for ingest, query, markdown edits, briefs, and evals
  • validation scripts and CI for index and fake-vault quality
  • governance rules for provenance, privacy, and publication
  • benchmark rubrics for retrieval, citation quality, synthesis, contradiction handling, and privacy

Attribution policy

Future documentation should preserve this order of credit:

  1. Andrej Karpathy for the LLM Knowledge Bases / LLM Wiki pattern.
  2. Garry Tan for GStack's agent-workflow discipline and GBrain's retrieval/brain architecture.
  3. This Memex repository for the packaged framework, templates, governance, evals, and public/private deployment model.

When describing the project externally, use this concise wording:

Memex is a local-first, agent-operable second-brain framework inspired first by Andrej Karpathy's LLM Knowledge Bases / LLM Wiki pattern, and operationally influenced by Garry Tan's GStack and GBrain.