PMB

Local-First Memory for Your AI Coding Agent

PMB provides a real memory for your AI coding tools like Claude Code, Cursor, Codex, and Zed. Instead of starting from scratch every session, PMB stores decisions, lessons, and facts in a local SQLite file on your disk. It is 100% on your machine—no API keys, no cloud, and no telemetry.

How it Works

PMB uses hooks to inject the right memory before the model thinks, and journals the agent's work after. This happens without an LLM call on the read path, meaning there is no tool the agent has to remember to call.

• Auto-recall: Every message is classified in sub-millisecond time; matching lessons, decisions, and project overviews are fetched before reasoning.
• Sub-millisecond Writes: The MCP tool returns instantly because SQLite writes happen first, while embeddings and vector inserts run on a background thread.
• Hybrid Recall: Fuses BM25 + dense vectors + entity graph with Reciprocal-Rank-Fusion (RRF) for high-quality retrieval.
• Honest Follow-Rate: Every rule is scored by whether the agent actually follows it. Useful ones get starred; ignored ones are flagged as dead so you can prune what doesn't help.

Why it Matters

Persistent memory changes your workflow by providing tangible outcomes rather than just features:

1. Stop re-explaining: Every session starts with the agent already knowing your decisions, conventions, and previous bugs.
2. Switch tools without losing context: Since Claude Code, Cursor, Codex, and Zed all read from the same memory, your context follows you across different editors.
bsp;
3. Trustworthy Memory: PMB scores whether each lesson gets followed, ensuring your context stays honest and not bloated with useless information.

The Dashboard

PMB includes a local web app dashboard that renders your project's memory in real-time.

• The Map: A live entity graph where every fact, decision, and lesson becomes a node. You can explore connections, hover to inspect details, and see the importance of color-coded nodes.
bsp;
• The Timeline: A git-graph style journal that automatically records every decision, lesson, and commit in chronological order.

Quickstart

Get up and running in 60 seconds with three simple commands:

1. Install: Run pip install pmb-ai to install the pure Python package on macOS, Linux, or Windows.
bsp;
2. Connect: Wire PMB into your agent using pmb connect claude-code (or other supported agents).
bsp;
3. Talk: Just work as usual. PMB records and recalls automatically. You can open the dashboard anytime with pmb dashboard.

Architecture & Performance

PMB is built on durable, local-first pieces. Every event lives in SQLite; vectors live in LanceDB next to it. It uses a hybrid router engine for recall and an async write embed queue to ensure zero blocking of the agent's turn.

• Recall Latency: ~35ms (warm daemon, cache off).
bsp;
• Write Speed: < 1ms async writes.
bsp;
• Accuracy: 94.6% recall@10 on LoCoMo benchmarks.
bsp;