the README already credits karpathy's gist as the foundational methodology. a few other repos have implemented the same pattern with different trade-offs — there's lucasastorian/llmwiki (claude MCP), Astro-Han/karpathy-llm-wiki (agent-skill), atomicmemory/llm-wiki-compiler (cli), and braindb, which i maintain.
a quick word on what braindb is, for context: braindb is a persistent memory database + REST API for LLM agents. It stores typed entities (thoughts, facts, sources, documents, behavioural rules) with explicit relations; recall combines fuzzy + semantic keyword search, graph traversal up to 3 hops with geometric-mean scoring, temporal decay, and always-on rule injection — so a query returns a ranked graph neighbourhood, not a pile of chunks. Two API surfaces: raw HTTP for any agent's tool-call layer, or a single /agent/query endpoint that handles recall + save + relate in plain English (built on the OpenAI Agents SDK + LiteLLM with pluggable providers). Two always-on sidecars do the hands-off work: a watcher auto-ingests anything dropped into data/sources/, and a wiki_scheduler runs a maintainer + writer agent loop that maintains a self-healing wiki on top of the graph — consolidating duplicates, revising pages when new evidence contradicts prior claims, citing every fact inline with [[ref:UUID]]. Postgres + pgvector. Inspired by Karpathy's LLM-wiki gist; goes further with structure, retrieval, and a real graph on top of the plain-markdown baseline.
would you accept a PR that adds a "Related implementations" or "See also" sub-section under Credits, seeded with those four? happy to write it neutrally — one line per repo, alphabetical, not framed as competition.
— @dimknaf
the README already credits karpathy's gist as the foundational methodology. a few other repos have implemented the same pattern with different trade-offs — there's lucasastorian/llmwiki (claude MCP), Astro-Han/karpathy-llm-wiki (agent-skill), atomicmemory/llm-wiki-compiler (cli), and braindb, which i maintain.
a quick word on what braindb is, for context: braindb is a persistent memory database + REST API for LLM agents. It stores typed entities (thoughts, facts, sources, documents, behavioural rules) with explicit relations; recall combines fuzzy + semantic keyword search, graph traversal up to 3 hops with geometric-mean scoring, temporal decay, and always-on rule injection — so a query returns a ranked graph neighbourhood, not a pile of chunks. Two API surfaces: raw HTTP for any agent's tool-call layer, or a single
/agent/queryendpoint that handles recall + save + relate in plain English (built on the OpenAI Agents SDK + LiteLLM with pluggable providers). Two always-on sidecars do the hands-off work: a watcher auto-ingests anything dropped intodata/sources/, and a wiki_scheduler runs a maintainer + writer agent loop that maintains a self-healing wiki on top of the graph — consolidating duplicates, revising pages when new evidence contradicts prior claims, citing every fact inline with[[ref:UUID]]. Postgres + pgvector. Inspired by Karpathy's LLM-wiki gist; goes further with structure, retrieval, and a real graph on top of the plain-markdown baseline.would you accept a PR that adds a "Related implementations" or "See also" sub-section under Credits, seeded with those four? happy to write it neutrally — one line per repo, alphabetical, not framed as competition.
— @dimknaf