PMB drops local-first memory for AI
PMB (Persistent Memory Bank) is a local-first, open-source memory system designed to solve the continuity problem for AI coding assistants like Claude Code and Cursor. Operating as an MCP server, it stores project context locally using SQLite and LanceDB, utilizing a hybrid retrieval pipeline to fetch memories in 35 milliseconds.
PMB represents a crucial paradigm shift in AI development by decoupling state and memory from specific coding assistants, ensuring a developer's context remains private, local, and portable.
* **Completely Local & Private:** All data resides in standard SQLite and LanceDB files on the user's disk, requiring no cloud telemetry, API keys, or external LLM calls.
* **Hybrid Recall Fusion:** Achieves high retrieval accuracy and sub-millisecond async writes by ranking combined BM25, vector, and entity graph outputs using Reciprocal Rank Fusion (RRF).
* **Self-Pruning Context:** Prevents context-window bloat by rating the follow-rate of stored lessons and automatically archiving or decaying ignored ones.
* **Interoperable Memory:** Since the memory lives on the local disk as an MCP server, developer context seamlessly transfers across different editors and agents.
DISCOVERED
1d ago
2026-06-28
PUBLISHED
1d ago
2026-06-28
RELEVANCE
AUTHOR
Oleksii Bondar