MCP Memory Server in · shared project memory — local-first · offline-ready

MCP Memory Server in Tokyo

Stop running a toy JSON memory server. MemHub is a production MCP memory server on PostgreSQL + pgvector — isolated per org/project, audited, and ready for your whole team in one command.

Most MCP servers die on one laptop. MemHub runs the same MCP protocol on a real database with row-level security — one endpoint, one credential per project, every agent in sync.

Start free — get briefed in 2 min

Free for 1 project · No card · Self-host or we host · No external API

acme/mcp-memory-server
synchronized

context/main

memories 41

pinned 6

contributors 9

BRIEFING FOR NEW SESSION

What changed since your last pull

3 new
01
decision

Standardize on idempotent write endpoints

Source: architecture.md

02
pattern

Repository layer owns every database transaction

Source: pull request #214

03
bugfix

Refresh tokens rotate atomically under load

Source: auth/session.ts

Claude Code
Cursor
Windsurf

WHY IT MATTERS HERE

Why Tokyo engineering teams choose MemHub

In Tokyo, software teams ship faster when their tools share one memory instead of three silos. Mixed toolchains stop fragmenting knowledge across vendor-specific stores.

Every entry carries authorship and version history, which keeps both humans and auditors comfortable.

  • Built for Tokyo: enterprise modernization programs at national scale.
  • Local-first writes keep sessions productive through outages and travel.
  • One command onboards the whole team: memhub login, then connect for Claude Code, Cursor, Windsurf, Cline, Codex CLI, or any JSON client.

UNIVERSAL COMPATIBILITY

Plugs into any CLI or development tool you use.

If it speaks MCP or reads a JSON config, MemHub plugs in. One command wires the majors; everything else joins as a generic client.

  • Claude Code
  • Cursor
  • Windsurf
  • Cline
  • Codex CLI (GPT)
  • GitHub Copilot
  • Google Antigravity
  • Gemini CLI
  • Aider
  • OpenCode
  • CI bots via REST

WHAT YOU GET

MCP Memory Server in Tokyo, without the trade-offs.

01

One proxy, every client

memhub serve bridges any MCP client over stdio. Zero daemons on the laptop, zero workflow change.

02

Queue when offline

Local SQLite outbox keeps memory writable during outages — syncs automatically when you’re back.

03

Prove who did what

Actor, action, and target on every admin change — ready for security review and audits.

04

No pasted secrets

mh login via browser with scoped, revocable credentials — a lost laptop loses access instantly.

05

Isolated by default

Row-level security per org/project — a bug in one repo can’t read another’s memory. Safe for the whole team.

06

Run on PostgreSQL you trust

pgvector-ready schema, versioned migrations, Docker/Easypanel. Backups with pg_dump — no exotic stack.

The honest comparison.

No feature-matrix gymnastics — just what changes the week you adopt MemHub.

Without MemHubWith MemHub
Context lives in heads and DMs → onboarding takes daysStructured, searchable memory → new dev/agent productive in minutes
Every agent re-learns the repo from scratchSessions open pre-briefed with only what changed — zero re-explaining
Rules files cap out at ~2k tokensUnbounded versioned history stays one query away — never cap again
Decisions re-litigated, bugs re-introducedPinned invariants prevent regressions — agents respect past calls

Keep your stack. Add shared memory.

One command for Claude Code, Cursor, Codex, Windsurf, Cline or any MCP client. No new editor. Setup in 2 minutes, works offline.

repository terminal

$ npm install --global @simplexlat/memhub

$ memhub login

✓ project resolved · context ready · memory synchronized

Questions, answered straight.

Can I self-host the MCP memory server?

Yes. The repository includes Docker Compose and Easypanel templates for PostgreSQL 16 with pgvector, plus migration and seed commands.

Which clients can connect?

Any MCP-compatible client: Claude Code, Cursor, Codex CLI, Gemini-based IDEs such as Antigravity, Windsurf, Cline, or a generic JSON config.

What happens when our connection drops?

Nothing interrupts the work. Writes commit to the local SQLite outbox and synchronize automatically when the shared service is reachable again.

Can we try it without sending data anywhere?

Yes — the local engine works fully offline, and the self-hosted stack keeps even the shared layer on your own PostgreSQL.

How much does it cost?

A Community tier is free, and the Team plan adds multi-project contexts, seats, and higher synchronization limits. Self-hosting is unlimited.

MORE LOCATIONS

EXPLORE OTHER USE CASES

THE NEXT SESSION STARTS HERE — 2 MIN SETUP

Stop rebuilding context.
Start shipping faster.

Create workspace — free