vibedonaldsvibedonalds.com
PMB screenshot

PMB

Free, open-source local memory for AI coding agents: one SQLite file on your disk, served to them over MCP.

What is PMB, in two sentences

PMB is a local-first memory layer for AI coding agents: decisions, lessons and project facts live in one SQLite file on your disk and come back to Claude Code, Cursor, Codex, Zed and other agents through MCP, with no cloud, account or API keys. It is free and open source under Apache-2.0, with no paid tier.

About PMB

You install it with pip install pmb-ai, or with npx pmb-ai setup, which installs the Python package and then runs setup. The pmb connect command writes an agent's MCP config and appends usage rules to CLAUDE.md or AGENTS.md; the MCP server then runs over stdio as a child process of the agent, with no network, port or token. Several agents can point at one shared workspace, and sharing a memory across machines or a team uses an optional HTTP mode with bearer-token auth.

How much happens automatically depends on the agent. Claude Code gets full hooks: memory is recalled before each message and written as the agent works, so memory works without the agent calling a tool. Codex gets MCP plus an ambient write when a turn completes but no per-turn read hook, and Cursor, Windsurf, Gemini, VS Code, Zed, OpenCode and Continue use memory through MCP tools.

Recall is hybrid, fusing BM25, dense vectors and an entity graph, with optional reranking, through Reciprocal-Rank-Fusion, and PMB makes no LLM call on the read path. Each lesson is scored by whether the agent actually follows it: useful ones are starred and ignored ones flagged dead, so you can prune them. The pmb dashboard command opens a local web app with a graph of your memory and a timeline of decisions, lessons and commits.

Vectors live in LanceDB next to the SQLite file, and the files can be copied anywhere with cp. Embeddings from MiniLM-L12-v2 cover 50+ languages, and the optional summarize and graph-extract passes run on whatever model you point them at, including a local Ollama. PMB is pure Python and says it is tested on Windows 11, macOS 14 and Ubuntu 22.04.

Sources: pmbai.dev, docs.pmbai.dev/introduction, docs.pmbai.dev/quickstart and docs.pmbai.dev/guides/connect, read 2026-10-27

PMB pros and cons

Pros

  • Memory stays in local files: PMB says there are no network calls on the read path, no account and no telemetry, and that it still works with the internet unplugged.
  • Claude Code, Cursor, Codex and Zed can read the same memory, so switching agents keeps your context.
  • Memory decays, archives and dedupes on its own, and PMB says it never deletes anything behind your back.
  • Disconnecting an agent leaves the stored memory on disk, still readable through the CLI.

Cons

  • Automatic recall before each message works only with Claude Code; Codex has no per-turn read hook, and the other agents reach memory only through MCP tools.
  • Connecting an agent edits its MCP config and appends rules to CLAUDE.md or AGENTS.md, and the agent has to be restarted.
  • It is a Python package you set up and inspect from the terminal or a local dashboard; there is no cloud version.

Tagged

  • Web-based

PMB alternatives

The closest alternatives to PMB.

  1. 01
    pumaDB

    a small hosted memory layer for AI agents

    →
  2. 02
    Pieces

    AI dev companion for snippets, screenshots of code, and conversational context across IDEs and browsers.

    Freemium →

Related reading

Reviews
No reviews yet — be the first to rate PMB.

Featured on Vibedonalds

Own PMB? Add this badge to your site to show you’re listed — and link back to your profile here.

PMB — Featured on Vibedonalds
<a href="https://vibedonalds.com/tools/pmb" target="_blank" rel="noopener nofollow">
  <img src="https://vibedonalds.com/badge/featured-on-vibedonalds.svg" alt="PMB — Featured on Vibedonalds" width="240" height="60" loading="lazy" />
</a>

Frequently asked questions

What is PMB?
PMB is a free, open-source local memory for AI coding agents: one SQLite file on your disk, served to them over MCP.
What platforms does PMB support?
PMB's own site lists it for CLI, macOS, Linux and Windows.
What are the alternatives to PMB?
The closest alternatives to PMB are pumaDB, Pieces.
What are the downsides of PMB?
Automatic recall before each message works only with Claude Code; Codex has no per-turn read hook, and the other agents reach memory only through MCP tools. Connecting an agent edits its MCP config and appends rules to CLAUDE.md or AGENTS.md, and the agent has to be restarted. It is a Python package you set up and inspect from the terminal or a local dashboard; there is no cloud version.
Is PMB free?
Yes: PMB is open source under Apache-2.0, and its site says it is free forever, with no paid tier, no seats and no telemetry.
Which agents does PMB work with?
PMB says any MCP-aware agent works, and pmb connect can wire Claude Code, Cursor, Codex, Windsurf, Gemini, VS Code, Zed, OpenCode and Continue. Claude Code gets full hooks with automatic recall, Codex gets ambient writes, and the others use memory through MCP tools.
Does PMB send my code anywhere?
PMB says no: memory lives in a local SQLite file with vectors in LanceDB next to it, with no network calls on the read path, no account and no telemetry. The optional summarize and graph-extract passes run on the model you point them at, which can be a local Ollama.
Which operating systems does PMB run on?
PMB is pure Python and says it is tested on macOS, Linux and Windows; the site names Windows 11, macOS 14 and Ubuntu 22.04. You install it with pip install pmb-ai or npx pmb-ai setup.
How is PMB different from RAG or a vector database?
PMB names two differences: recall is hybrid, combining BM25, dense vectors and an entity graph, and it is automatic, with the right memory injected before the model thinks. That automatic injection comes from the full hooks, which the docs list for Claude Code.