LLM Wiki vs OpenMemory vs GBrain: which should I use?

As of November 6, 2026: use the LLM Wiki pattern (Karpathy's April 2026 gist) when you want an agent to compile your sources into plain markdown pages you can read in Obsidian, with no database; use GBrain (garrytan/gbrain, MIT) when you want that knowledge searchable by several agents over MCP, with sources, corrections and optional semantic search; and treat 'OpenMemory' carefully, because the name now covers different projects: Mem0 removed its OpenMemory MCP server from the mem0 repository on July 29, 2026, mem0ai/openmemory is now a CLI that moves coding sessions between Claude Code, Codex and OpenCode, and CaviraOSS's OpenMemory was renamed LongMemory.

What is the difference between an LLM wiki and a memory layer?

An LLM wiki is a set of markdown pages an agent writes and keeps current: when you add a source, the agent updates entity and topic pages, so knowledge is compiled once and read later. A memory layer such as Mem0 or GBrain stores facts in a database and retrieves them per query, with search, scoping and access control. The wiki is easier to read and audit; the memory layer scales better and serves many agents at once.

What happened to OpenMemory MCP?

Mem0 announced OpenMemory MCP as a local, private memory server for MCP clients, but on July 29, 2026 it removed OpenMemory from the mem0 monorepo (commit 'remove OpenMemory from the monorepo'). The name now belongs to mem0ai/openmemory, a beta CLI and TUI that imports and exports sessions between Claude Code, Codex and OpenCode. For agent memory, Mem0 itself (Apache-2.0, free hosted plan) is the maintained product.

Yes. GBrain is MIT-licensed and free to self-host; gbrain.io is optional paid hosting. You can start keyless with keyword retrieval in a local PGLite database, then add an embedding provider (Voyage voyage-4 is the default for new installs) and a reranker when you want semantic search. Model and embedding API calls are billed by those providers.

Quick Answer

LLM Wiki vs OpenMemory vs GBrain: Which AI Memory? (2026)

Published: November 6, 2026

The short answer

They are three different layers, not three versions of one tool. As of November 6, 2026:

LLM Wiki — a pattern: an agent writes and maintains an interlinked markdown wiki from your sources. Best for one person’s research, reading notes or project knowledge you want to read yourself.
GBrain — a system: a memory layer with sources, corrections and search that several agents share over MCP. Best when the knowledge base outgrows what an agent can read from an index file.
OpenMemory — a name, now attached to different projects. Mem0’s OpenMemory MCP server is gone from Mem0’s repository; check which “OpenMemory” a guide means before you install it.

The comparison

	LLM Wiki (Karpathy pattern)	GBrain	”OpenMemory” today
What it is	Idea file: agent maintains markdown wiki	Memory layer + agent skills	mem0ai/openmemory: session-porting CLI · CaviraOSS project renamed LongMemory
Storage	Markdown files (often in Obsidian)	PGLite by default; Postgres + pgvector for shared or large brains	Your harness’s session files · LongMemory: SQLite
Retrieval	Agent reads `index.md`, then pages	Keyword (keyless) → hybrid vector search + reranker	n/a · LongMemory: MCP + HTTP recall
Multi-agent access	Whatever can read the folder	MCP (Claude Code, Codex, Cursor, Windsurf…), remote over Tailscale	n/a · LongMemory: MCP
Licence / cost	Free, no software	MIT, free; gbrain.io hosting optional	MIT · LongMemory Apache-2.0
Activity (Nov 6, 2026)	Pattern; many implementations	v0.60.589.0 released Nov 6; ~31,500 stars	openmemory last push Jul 29 · LongMemory last push Oct 18

Repository data from GitHub on November 6, 2026.

LLM Wiki — compile knowledge once

Andrej Karpathy published the llm-wiki gist in early April 2026 as an “idea file” to paste into an agent such as Claude Code or Codex. His argument: in RAG “the LLM is rediscovering knowledge from scratch on every question”; instead, the agent reads each new source and integrates it into persistent pages, updating entity pages, revising summaries and flagging contradictions. An index.md lists pages and a log.md records ingests and queries. Karpathy says it “works surprisingly well at moderate scale (~100 sources, ~hundreds of pages)” without embedding infrastructure. He browses the result in Obsidian: “Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.”

Choose it when one person (or a small team) wants readable, auditable knowledge and a few hundred pages is enough. Background: what is LLMWiki.

Where it stops: past a few hundred pages the agent cannot read the index cheaply, nothing enforces who may read what, and two agents editing the same pages collide.

GBrain — the wiki idea with a database behind it

GBrain (garrytan/gbrain, MIT, created April 5, 2026) is Garry Tan’s production setup for his OpenClaw and Hermes agents; the README reports 155,795 pages, 24,589 people and 5,340 companies in his own brain. It “stores explicit facts with their sources, supports corrections and withdrawal, and makes the same memory available across your agents.” Key mechanics:

Start keyless: gbrain init --pglite --no-embedding gives a local brain with keyword search and no model API or server.
Upgrade search when needed: embeddings from Voyage (default voyage-4), OpenAI, Gemini, Ollama, LM Studio and others; a Voyage reranker by default in the balanced modes; Postgres + pgvector for shared or large brains.
Connect agents: gbrain setup claude-code or gbrain setup codex registers the MCP server and adds “brain-first” instructions; gbrain mcp expose publishes it on your Tailscale tailnet for remote and cloud agents.
Company brain: remote clients are limited by source and operation grants plus visibility filters; the README warns local files and shared database credentials are a separate trust boundary.

Choose it when several agents should share one memory, you want citations behind every answer, or the wiki has outgrown an index file.

OpenMemory — check which one you mean

Assistants still answer “use OpenMemory MCP” from 2025 posts. The current state:

Mem0’s OpenMemory MCP — announced as a local, private MCP memory server; removed from the mem0 monorepo on July 29, 2026. Mem0 itself remains maintained (Apache-2.0, ~68,500 stars) with a free hosted plan: unlimited end users, 10,000 adds and 1,000 retrievals a month for 30 days on Mem0’s model calls, then 50,000 adds and 10,000 retrievals with your own LLM key.
mem0ai/openmemory — now a beta CLI/TUI that imports and exports coding sessions between Claude Code, Codex and OpenCode. Useful when switching harnesses; not a knowledge base.
CaviraOSS OpenMemory → LongMemory — the GitHub repository now redirects to CaviraOSS/LongMemory: a local-first, SQLite-backed memory engine with provenance, temporal recall and an MCP server (Apache-2.0).

Choose Mem0 if you wanted OpenMemory for chatbot or assistant memory; LongMemory if you wanted a local MCP memory server.

How to choose

Is it your own reading and research, under a few hundred sources? LLM Wiki in a git repo or Obsidian vault.
Do several agents need the same knowledge, with sources? GBrain.
Do you need per-user memory inside a product? Mem0 or another memory layer: see best AI agent memory systems.
Just carrying context between coding agents? mem0ai/openmemory, or a session-memory plugin such as claude-mem.

The patterns combine well: GBrain’s source of truth can be the same markdown a wiki agent writes. For the broader category, see knowledge layer for large language models.

Last verified: November 6, 2026.