PMB Ai

PMB is a local-first memory system for AI coding agents. It stores decisions, lessons, and project facts in a single SQLite file on the user's machine to prevent context loss between sessions. The system feeds this memory back to agents through the MCP protocol.

It works with Claude Code, Cursor, Codex, and Zed. Memory is automatically surfaced before the agent generates a response, with every message classified in sub-millisecond time and relevant lessons, decisions, and project overviews fetched in 4-16 milliseconds. Writes occur asynchronously in under one millisecond using SQLite for primary storage and LanceDB for vectors. Hybrid recall combines BM25 ranking, dense vectors, an entity graph, and optional reranking fused by Reciprocal-Rank-Fusion, achieving 94.5 percent recall at ten results. Lessons receive a follow-rate score that identifies useful items for starring and flags ignored ones for pruning.

The included local web dashboard presents memory as an interactive entity graph with nodes for facts, decisions, lessons, files, and entities that are color-coded by type and sized by importance. A git-style timeline journal displays events by project. All operations remain offline with no API keys, cloud services, telemetry, or network dependencies. The tool installs via a single pip command, runs on macOS, Linux, and Windows, and supports both local and hosted models including Ollama for any summarization or graph extraction steps. Data resides entirely on disk and can be inspected or copied directly.

PMB is delivered as an open-source Apache 2.0 project. Installation, connection to an agent, and dashboard access each require one command with no accounts or external configuration needed. It operates as a child process over stdio with no ports or tokens.