Impact-Site-Verification: 41b53a0c-6d04-458b-a457-fe9e29acde1a

Developer Tools/UnknownOpen SourceDeveloper ToolsArtificial Intelligence·

PMB

Local-first persistent memory for AI coding agents

PMB

What it does

PMB is an open-source, local-first memory layer for AI coding agents like Claude Code, Cursor, Codex, and Zed. It stores decisions, lessons, project facts, and recent work in a single SQLite file on your disk. Through the Model Context Protocol (MCP), PMB automatically injects relevant memories before the agent responds and journals the agent's actions afterward—all without any cloud dependency, API keys, or LLM calls on the read path. The system uses hybrid recall (BM25 + dense vectors + entity graph) to surface the most useful memories in ~35 ms, and it includes a local dashboard with a map and timeline for inspecting memory.

Who it is for

PMB is designed for developers who use AI coding agents and are frustrated by the lack of persistent context between sessions. It targets individual developers and small teams working with Claude Code, Cursor, Codex, Zed, or any MCP-aware agent. The tool is especially useful for those who value privacy, offline capability, and full control over their data—since everything runs locally with no telemetry or network calls on the read path.

Why it matters

Without persistent memory, every new session with an AI coding agent starts from scratch. Developers must repeatedly re-explain project conventions, past decisions, and ongoing bugs. PMB eliminates this friction by making memory automatic and trustworthy. It scores each memory by how often the agent actually follows it, flagging unhelpful or ignored rules so the context stays lean and relevant. Over time, memory compounds: the agent gets smarter with each session, and switching between tools (e.g., from Claude Code to Cursor) costs nothing because they all read the same local memory.

Launch signal

PMB launched with a clear, well-documented website, a GitHub repository (420 stars at the time of writing), and a PyPI package. The project is Apache 2.0 licensed and emphasizes radical honesty—it tracks which memories are actually used and flags dead ones. The website includes benchmarks (94.6% recall@10 on LoCoMo, p50 recall latency 24 ms for ~100-memory workspaces) and a detailed architecture section. The founder, Oleksii Bondar, appears to be the sole creator based on available evidence.

Brand and naming

The name "PMB" is short but cryptic—it doesn't convey the product's purpose, which may hinder recall and discovery. The tagline "local-first memory for your AI coding agent" is descriptive but not particularly catchy. The brand positioning as "memory for AI agents" is clear and addresses a real pain point, but it competes in a crowded space of similar tools (e.g., memory for LLMs). The .dev domain and open-source Apache 2.0 license reinforce the developer-focused, trust-oriented positioning.

Founder

Oleksii Bondar

Get more like this in our weekly newsletter.