Tutorial demonstrates how to add GBrain's markdown‑first self‑wiring Memory Layer to AI agents

A step‑by‑step coding tutorial published May 22, 2026 shows how to install and run GBrain v0.38.2.0, the open‑source memory layer Garry Tan developed to power his OpenClaw and Hermes deployments. The walkthrough reproduces a live install from Bun setup to a working brain, captures terminal output for each command, and claims a roughly 20‑minute end‑to‑end experience for the small‑demo path. For builders, the guide provides a reproducible recipe to add persistent, self‑wiring memory to agent stacks.

GBrain is presented as a markdown‑first, Postgres‑backed knowledge layer that auto‑wires a typed knowledge graph using regex inference instead of relying on LLM calls for entity wiring. The tutorial uses PGLite (embedded Postgres 17 compiled to WASM) to provision a local ~/.gbrain/brain.pglite with no server or Docker required. The codebase ships on Bun; the canonical install command shown is bun install -g github:garrytan/gbrain, and the repository is MIT‑licensed at github.com/garrytan/gbrain.

Feature coverage in the tutorial includes a hybrid search CLI that combines vector search, BM25 keyword scoring, and Reciprocal Rank Fusion (RRF), with a ZeroEntropy reranker set as the default. The walkthrough notes the scale of the production brain backing Garry’s agents: 146,646 pages, 24,585 people, 5,339 companies and 66 autonomous cron jobs. An MCP server exposes 74 tools so models such as Claude Code, Cursor and Windsurf can read from and write to the brain directly.

The guide also documents practical integration constraints builders must plan for: embeddings are optional for the install path but required to produce vector search results; ZeroEntropy is the default embeddings provider and OpenAI or Voyage keys are supported, while Anthropic keys are optional for multi‑query expansion. Without an embedding API key, gbrain will still run keyword‑only queries. The tutorial captures schema migration output during initialization (schema v1→v85) to show real‑world setup behavior.

For practitioners, the tutorial highlights concrete implications: PGLite lets teams run an embedded Postgres instance locally in seconds, the regex‑based graph extraction avoids LLM costs for entity wiring, and the hybrid search stack plus MCP toolset provide a clear path to connect agents to persistent memory. The walkthrough lists prerequisites (macOS/Linux or WSL2, a code editor, Bun ≥1.3.10) and reproduces the exact init commands used in the live run so developers can follow along step by step.