Local-first memory for coding agents

PMB gives coding agents local memory

PMB gives Claude Code, Codex, Cursor, and other MCP agents a shared private memory. It captures project decisions and lessons as you work, ranks relevant context locally, and injects it before the next answer.

Start with setup See how it works

Local by default SQLite, LanceDB, and config live on your machine.

Fast recall One warm daemon serves every connected agent.

Agent-native MCP tools plus hooks fit into existing workflows.

How PMB fits into your workflow

flowchart LR
  Work["Work"] --> Capture["Capture"]
  Capture --> Store["Store"]
  Store --> Recall["Recall"]
  Recall --> Context["Context"]
  Context --> Answer["Answer"]
  Answer --> Work

Choose your path

• Getting started

  Go from zero to a wired agent with pip, uv, pipx, or npx.

  Install PMB →

• Usage

  Connect Claude Code, Codex, Cursor, Windsurf, Gemini, Zed, and more.

  Connect an agent →

• Architecture

  See the daemon, hook, MCP, storage, and retrieval paths as diagrams.

  Understand the system →

• Core engine

  Read the Engine map, schema, queues, and code paths behind recall.

  Inspect the core →

• Commands

  Look up setup, recall, dashboard, delete, sync, and maintenance commands.

  Open the reference →

Why teams use it

Durable memory Facts, decisions, goals, lessons, and session summaries persist between sessions.

Free reads Local pre-message recall through hooks and MCP, so remembering costs no tokens.

One shared brain Several agents on one project share a warm daemon and one workspace memory.

Private by default Local storage; the network is touched only by the explicit sync commands you run.

Learns from drift Lessons and failures surface before the work they apply to, so mistakes do not repeat.

The core loop

flowchart TB
  User["Request"] --> Prepare["Prepare"]
  Prepare --> Context{"Found?"}
  Context -->|Yes| Relevant["Relevant memory"]
  Context -->|No| Quiet["Stay quiet"]
  Relevant --> Agent["Act"]
  Quiet --> Agent
  Agent --> Record{"Worth saving?"}
  Record -->|Yes| Memory["Record"]
  Record -->|No| Skip["Skip"]
  Memory --> Next["Better next turn"]
  Skip --> Next

Common tasks

I want to...	Do this
Set up one agent	pmb setup
Set up every detected agent	pmb setup --all
Install from npm in one step	npx pmb-ai setup
See whether memory is warm	pmb daemon status
Change the embedding model	pmb model
Archive a memory	pmb delete <ulid>
Delete a memory permanently	pmb delete <ulid> --hard
Bring an archived memory back	pmb restore <ulid>
Open the local dashboard	pmb dashboard
Reset stray PMB processes	pmb daemon kill-all

What to read next

• Guide: install PMB, connect agents, and operate the dashboard.
• Concepts: understand the architecture and retrieval model.
• Reference: look up commands and configuration.
• Contributing: extend language packs or improve the docs.