MemBukkit – LLM memory that can answer "what was true in May?" (github.com)

MemBukkit has introduced a novel long-term memory system for LLM applications that enables users to track and retrieve dated atomic facts from conversations and documents. Significantly, when a fact changes (such as job status or rent), the previous information is marked as superseded rather than deleted, allowing users to query historical truths, like "what was true in May?" accompanied by detailed receipts that document the evidence and costs associated with the queries. MemBukkit achieved a state-of-the-art score of 92.6% on the LongMemEval-S benchmark, outperforming existing systems while significantly reducing the amount of data read by optimizing memory access. The architecture uses a two-lane indexing strategy, separating dated facts from verbatim source turns, which enhances retrieval efficiency. This unique design allows for rapid index creation and avoids the rigid structures inherent in traditional approaches. MemBukkit can be deployed locally and configured easily through various methods, including Python and Docker, ensuring user data privacy while offering the flexibility of using both local and hosted models. This innovative approach aims to improve the quality of context-aware responses and historical queries, setting a new standard for memory management in AI and ML applications.