PMB Review (2026): Pricing, Features & Honest Verdict

PMB Review: The Verdict

PMB is a local-first memory layer for AI coding agents that stops them forgetting your project every session. It runs as an MCP server on your machine, stores decisions and lessons in SQLite, and injects relevant context before the model responds. Best for developers who use Claude Code, Cursor, or Codex daily and are tired of re-explaining their codebase. Price: free and open source. Rating: 8/10.

Every AI coding agent has the same flaw. It is brilliant for an hour, then the session ends and it forgets everything. The architecture decision you explained, the bug that bit you last week, the convention you insisted on twice already, all gone. PMB, an open-source tool from Oleksii Bondar, fixes that by giving your agent a memory that lives on your disk. No cloud, no API keys, no re-explaining the same project every morning.

What is PMB?

PMB is persistent memory for AI coding agents. It stores your project’s decisions, lessons, goals, recent work, and documentation in a local SQLite database, then retrieves the relevant pieces and feeds them to your agent before each prompt. It works through MCP, the Model Context Protocol, which means it plugs into any MCP-aware agent: Claude Code, Cursor, Codex, Zed, and Windsurf are all supported. The whole thing runs offline on your machine.

The tagline says it plainly: your AI agent forgets every session, PMB gives it memory, on your disk. It launched in 2026 and is Apache-2.0 licensed, so it is genuinely free and the code is on GitHub for anyone to inspect. What makes it interesting is not just that it remembers, but how carefully it handles the remembering. There is no LLM call on the read path, no telemetry, and no external service holding your project’s context hostage. It is a small, focused tool that does one annoying thing well.

What Are PMB’s Key Features?

Automatic memory injection

You do not have to tell the agent to check its memory. PMB surfaces relevant context before the model responds, with no manual tool calls. So the agent starts each turn already knowing the decisions and lessons that matter, instead of waiting for you to remind it or paste in a context dump. This is the difference between memory that helps and memory you have to babysit.

Hybrid recall

Retrieval is where most memory tools fall apart, and PMB takes it seriously. It combines BM25 keyword search, dense vector embeddings, and entity graphs, and claims 94.5% recall@10 with results in around 35 milliseconds. The blend matters: keyword search catches exact terms, vectors catch meaning, and the entity graph ties related things together. Fast and accurate retrieval is the whole game, and the numbers here are strong.

Lessons that self-evaluate

This is the clever part. PMB scores whether the agent actually follows the rules you store. If a lesson keeps getting ignored, it gets flagged as “dead” so you can prune it. So your memory does not silently rot into a pile of rules nobody honors. It is a feedback loop on the memory itself, which is something almost no other tool in this space does.

Local-first architecture

Everything runs on your machine in a single SQLite file, with vectors stored in LanceDB. No API keys, no cloud, no telemetry. The MCP server talks to your agent over stdio, and embeddings are written asynchronously so they never block your turn. For anyone nervous about pumping their proprietary codebase context into someone else’s servers, this is the right design.

How Much Does PMB Cost?

PMB is free. It is Apache-2.0 open source, you install it with pip install pmb-ai, and there is no paid tier, no subscription, and no cloud bill. The repository is public on GitHub. The only cost is the few minutes of setup and the small amount of disk the SQLite file uses.

Against the alternatives, free and local is a real edge. Hosted memory services for AI agents exist and often charge per seat or per call, and they put your context on someone else’s infrastructure. PMB asks for none of that. If it saves you even one “let me re-explain the project” conversation a day, it has already paid for itself many times over, because it never cost anything to begin with.

Who is PMB Best For?

Use PMB if you code with an AI agent every day and you have hit the wall of re-explaining your project constantly. It is built for developers on Claude Code, Cursor, Codex, Zed, or Windsurf who want their agent to actually remember architecture decisions and hard-won lessons across sessions. The local-first design makes it especially good for anyone working on proprietary or sensitive codebases who cannot send context to a cloud service.

Skip PMB if you do not use an MCP-aware agent, since the whole thing rides on MCP. Skip it too if you only use AI for occasional one-off questions rather than sustained work on a real project, because the value compounds over long-running codebases and barely shows up for quick tasks. And if you are allergic to anything that needs a command-line install, the setup will not be for you, though it is a single pip command.

Best PMB Alternatives

Cursor’s built-in rules and memory handle some of this inside the editor. Convenient if you only use Cursor, but tied to that one tool and far less sophisticated on retrieval. PMB works across every MCP agent.

Mem0 is a popular memory layer for AI applications, with a hosted option. More general purpose and cloud-capable, but that also means API keys and sending context off your machine. PMB trades breadth for local-first privacy and a coding-specific focus.

A plain CLAUDE.md or rules file is the manual version: a markdown file the agent reads. Free and simple, but static, dumb, and growing it by hand gets old fast. PMB automates the retrieval and self-prunes dead rules instead.

Final Verdict: Is PMB Worth It?

PMB solves a real and constant annoyance for anyone who codes with AI agents. The forgetting problem is the single most frustrating thing about these tools, and PMB attacks it with a thoughtful design: automatic injection so you do not babysit it, hybrid recall that is fast and accurate, and self-evaluating lessons so your memory does not rot. That it is free, open source, and entirely local is the cherry on top.

The catches are minor and honest. It is a solo open-source project, so support is community-driven and the long-term roadmap depends on one maintainer. The recall numbers are the maker’s own benchmarks, not ours. And it only matters if you live in MCP-aware agents. But none of that changes the core fact: it costs nothing, it installs in one command, and it fixes something that bugs every AI coder daily. If you use Claude Code or Cursor seriously, install it this afternoon and see.