PMB
PMB is a local-first memory system designed for AI coding agents, enabling them to retain context across sessions. It stores project decisions, lessons, facts, and file references in a single SQLite file on the user's machine, paired with a LanceDB vector store for dense embeddings. All operations run offline, with no cloud dependencies, API keys, or telemetry, ensuring complete data privacy and control. The tool integrates seamlessly with popular AI-powered coding assistants—including Claude Code, Cursor, Codex, and Zed—via the Model Context Protocol (MCP). Every user prompt is classified in sub-milliseconds, and relevant memories are automatically injected into the agent's context before it responds. Writes happen asynchronously, returning instantly while the background thread embeds and indexes the data. PMB employs a hybrid recall mechanism combining BM25, dense vectors, and an entity graph, fused with Reciprocal-Rank-Fusion, achieving a recall@10 of 94.5%. It also scores each stored lesson by how often the agent actually follows it, flagging ignored or stale memories for pruning. This ensures that the agent's context remains tight and trustworthy. A local web dashboard provides an interactive graph visualization (the Map) and a timeline-based journal, allowing users to explore all stored entities, connections, and memory chunks. The entire system is open-source under Apache 2.0, installable via pip, and runs on macOS, Linux, and Windows.
- Category
- Apis
- Pricing
- Free
Problem solved
AI coding agents typically forget all context after each session, forcing developers to re-explain project conventions, past decisions, and known issues. PMB solves this by giving agents a persistent, local memory that automatically recalls relevant information before every prompt, eliminating repetitive context sharing and keeping projects on track.
Use cases
- Preserve project-specific conventions, decisions, and bug workarounds across AI coding sessions.
- Share a single memory store among multiple AI coding tools (Claude Code, Cursor, Codex, Zed) via MCP.
- Automatically inject relevant codebase context and lessons into agent prompts to improve code generation.
- Review and prune stale or unhelpful memories using follow-rate scoring.
- Explore project memory as an interactive graph and timeline for debugging or knowledge discovery.
Audience
- Developers
- ML Engineers
- Data Scientists
Pros
- Fully local and offline—no cloud, API keys, or telemetry.
- Open-source under Apache 2.0, free to use and modify.
- Fast hybrid recall (~35ms) with high accuracy (94.5% recall@10).
- MCP-native: works with Claude Code, Cursor, Codex, and Zed out of the box.
- Includes a rich visual dashboard for exploring memory graphs and timelines.
Cons
- Requires local Python environment and dependency installation.
- Currently limited to AI coding agents supporting MCP; not a general-purpose memory layer.
- May need tuning for very large codebases to maintain fast recall.
- Dashboard can become slow with thousands of connected entities.
- No built-in cloud sync or multi-user collaboration.
Official website: https://pmbai.dev/
Canonical LambdaBase page: https://www.lambdabase.com/tools/pmb