OpenBrain

#12Shared Multi-Agent Memory (Built by Us)

How we workBuilt by us

1. What we use it for: Built by us in TypeScript, Express, and PostgreSQL + pgvector, OpenBrain is the shared memory and MCP server that gives Claude Code, OpenAI Codex, and our scheduled agents one verified source of truth across every project.

2. Client use case · All client builds (RiseSync, a print and signage company, a members' dinner club, and a nonprofit education center)State: live

Problem

Across 947 human coding sessions and thousands of automated runs, isolated AI agents forget contract boundaries, repeat solved bugs, or act on stale project notes.

What the tool did

Built OpenBrain (3,534 commits) with hybrid semantic + keyword search (`ob`), freshness footers (`newer_count`), centralized postmortem lessons (`~/Work/openbrain/lessons/`), and a Dolt-backed Beads task tracker (`bd`).

Result

Queried via `ob` in 592 of 947 human coding sessions (and `bd` in 183 sessions) since 2026-07-24—including the cross-repo audit that verified every claim on this page.

3. What it costs us or the client

Built by us ($0 software license); runs on local PostgreSQL + pgvector and Docker over Tailscale with a few dollars per month in embedding tokens (as of 2026-10-22).

4. Setup notes (how it is wired in)

A Node 20 / TypeScript strict ESM server (`src/index.ts` and `src/mcp/server.ts`) backed by PostgreSQL + pgvector and Zod-validated MCP tools (`ob search-documents`, `ob get-context`), paired with a local Dolt server for Beads (`bd`) task tracking.

5. What broke & how it was handled

Early agent sessions dumped raw JSON documents containing 1,536-dimension embedding arrays straight into the prompt context (wasting tens of thousands of tokens) or trusted an older summary after a newer email or commit had changed the plan. We fixed both by stripping embedding vectors on read and adding a freshness footer (`newer_count > 0`) that requires agents to open newer records before answering.

6. What we compared it against & why we picked it

Compared against Notion MCP (used in only a handful of connector sessions), scattered per-repo Markdown notes, and hosted third-party memory APIs. We built OpenBrain ourselves so every harness we run—Claude Code, Codex, and Hermes—shares one local, auditable pgvector database with freshness enforcement.

7. Last reviewed: 2026-10-228. Video / tutorial: Episode 4 script package ready: "How We Run Claude Code + Codex + OpenBrain Across Client Builds"

All stack tools

Three separate steps. Stop after any of them.

Step 1 is a free 15-minute fit call. Step 2 is the AI Staff Assessment, $500, a written assessment you keep. Step 3 is your first hire at a published price, or you take the assessment and stop.

Step 1

Fit call

$0 · 15 min

tell me which job eats your week: the phone, the inbox, marketing or support
I tell you whether an AI hire fits, and whether I am the right person for it
you leave with the first hire I would start with
a slide deck

video call · same week

book it

Step 2

AI Staff Assessment

$500

a 60-minute video call about how your business runs
my written assessment the same day: whether an AI hire would pay for itself, the job I would hand off first, what to settle first, and the fixed price for that first hire
not taken off the build, and no separate quoting call: the assessment names the package at its published price
proposals written before anyone learned your business

paid when you book · not taken off the build

start here

Step 3

Your first hire

fixed price

Front Desk Hire or Brain + First Hire, at the published price
built in 2 to 6 weeks, then 30 days of tuning
or take the assessment and stop there
bench developers you never met

priced in the assessment

see the prices

Shared Multi-Agent Memory (Built by Us)OpenBrain

OpenBrain

Three separate steps. Stop after any of them.

Fit call

AI Staff Assessment

Your first hire

Want OpenBrain wired into your business?