PMB

PMB is a local-first memory layer designed for coding assistants that need persistent project history without sending data to remote servers. Every event, decision, and lesson is stored on disk using SQLite paired with vector embeddings in LanceDB, giving agents a reliable way to recall context across long-running sessions. Because all storage runs on the developer's machine, sensitive project knowledge stays private and under the user's direct control. Speed is a core part of the design. Reads typically resolve in tens of milliseconds, while write operations return in under one millisecond because embeddings and vector inserts run on a background thread. This low-latency behavior means agents can check memory between tool calls without slowing down the development loop. Teams running multiple agents benefit from consistent recall without paying for cloud storage or external APIs. The included Map and Timeline dashboard turns stored memory into something developers can actually inspect. A live entity graph reveals relationships between files, decisions, and lessons, while a git-style timeline surfaces commits, facts, and contextual changes over time. This visibility makes it straightforward to audit what an agent has learned, prune outdated entries, and keep the memory layer tidy. PMB also introduces honest lesson scoring, tracking whether the agent actually follows the rules it has recorded. Helpful lessons get starred, while ignored ones are flagged as dead so teams can remove misleading guidance. Combined with local storage, this scoring system prevents the bloat and drift that often plague long-lived agent memory setups.