membukkit

MemBukkit is long-term memory for LLM apps that shows its work. It turns conversations and documents into dated atomic facts. When something changes (new rent, new job), the old fact is superseded, not deleted, so you can ask "what was true in May?" and get May's answer. Every answer ships with receipts: the facts used, where they came from, what the query cost. It sets the state of the art on LongMemEval-S under the benchmark's official judge, 92.6%, while reading a small fraction of the tokens full-context reading pays. Python, CLI, local GUI, HTTP, and MCP, all over the same stores. Apache-2.0.