LLBBL Blog

LLBBL Blog

Amnesia-Free Agents: GBrain and the LLM Wiki Idea

Every agent I work with has amnesia.

Open a new terminal session, spin up a subagent, start a fresh chat, and the thing forgets everything. Your project guidelines, your coding style, your database schemas, the meeting where you decided why the schema looks like that. Gone. So you do one of two things: stuff a few thousand tokens into a bloated system prompt, or hand-maintain fragile context files like CLAUDE.md. I’ve done both. Neither feels good.

So when Y Combinator CEO Garry Tan open-sourced GBrain on April 5, 2026, a markdown-first memory layer, I wanted to understand what it does and whether it’s worth the setup.

What GBrain actually is

GBrain is a Postgres-native knowledge brain that acts as a persistent retrieval layer for your coding agents. The interesting part isn’t that it stores stuff, it’s how opinionated it is about the shape of that storage.

Three ideas do the heavy lifting:

  • Compiled truth vs. an append-only timeline. Raw daily logs, Slack messages, and emails go in one bucket. The curated “current facts” (contacts, active projects, decisions) live as clean Markdown in another. One is history, the other is truth.
  • A self-wiring knowledge graph. It parses your plain Markdown and pulls out relationships (works_at, invested_in, founded) using local grammars. No LLM API calls to build the map.
  • The Dream Cycle. An overnight cron job that crawls your daily logs, reconciles contradictions, checks citations, and rebuilds the search index while you sleep.

That last one worth a callout: Memory that maintains itself.

The category error everyone makes

People keep comparing GBrain to runtime agents like OpenClaw or Hermes, and that comparison doesn’t hold up. They’re different layers entirely.

GBrain isn’t an execution engine. It’s closer to a self-writing wiki.

  ┌────────────────────────────────────────────────┐
  │                AGENT RUNTIMES                    │
  │       (OpenClaw, Hermes, Claude Code)            │
  │   runs code, spawns subagents, hits the shell    │
  └─────────────────────────┬────────────────────────┘
                            │  reads & writes via MCP
                            ▼
  ┌────────────────────────────────────────────────┐
  │                  THE LLM WIKI                    │
  │               (GBrain, OpenWiki)                 │
  │    compiles raw history into clean markdown      │
  └────────────────────────────────────────────────┘

The runtime is the hands and eyes. It sandboxes code, runs terminal commands, touches the file system. The wiki is the long-term memory. You wire them together over the Model Context Protocol (MCP), and when the agent needs context, it doesn’t parse your raw email history. It queries the wiki and gets back synthesized, cited facts.

GBrain isn’t the only thing playing this role. OpenWiki, from the LangChain folks, uses the same trick to compile big source repos into navigable markdown. Same pattern, different input.

The payoff: your agent stays lightweight and task-focused, and your knowledge lives in a clean, git-versioned wiki instead of leaking into a system prompt.

What people actually run into

The concept is great. The reality has some sharp edges, and the community forums are honest about them.

It needs babysitting. GBrain moves fast, which means frequent schema updates and migrations. You’ll find yourself running things like gbrain doctor --remediate to fix database drift. It is not set-it-and-forget-it.

Tool sprawl adds up. To feed the Dream Cycle good data, people bolt on voice transcribers, contact enrichers, email scrapers. Most of those are paid SaaS. String enough together and you’ve built yourself a subscription problem.

Setup is a project. A Dockerized Postgres with pgvector, system keyring wrangling (macOS Keychain or D-Bus), cron schedulers. Budget anywhere from 30 minutes to a lost afternoon.

Should you bother?

If you’re juggling complex work across a bunch of codebases, and you want an agent that knows your past architecture decisions and updates its own docs overnight, the GBrain + OpenClaw setup earns its keep. That’s a real moat.

If you just want a lightweight assistant that remembers a few things, you’re better off with plain MEMORY.md files or a hosted API like Mem0 until the local graph-RAG tooling settles down. I’m somewhere in the middle right now, so I’ll watching it to see where it goes.

Amnesia is the default. It doesn’t have to be.

Sources

I’d appreciate a follow. You can subscribe with your email below. The emails go out once a week, or you can find me on Mastodon at @[email protected].

AI Developer-tools Agents LLM Memory