what is nexsandglass-agent-dedicatedmemory fr?

NexSandglass is a local memory engine for AI agents that goes beyond storing what you said, tracking who you are and how you are changing over time. It runs entirely on your machine using standard Python libraries and SQLite, with no API key needed and no external dependencies. You can plug it into an AI assistant as an MCP server and it will give that assistant a persistent, evolving understanding of you across sessions. Most memory systems for AI act as filing cabinets: they store conversations and retrieve relevant chunks when needed. NexSandglass works differently. It tracks a concept called Drift Velocity, which measures how your decisions shift over time. For example, if you consistently pick budget options and then start picking more expensive ones, that shift is recorded as a drift. The system also builds a knowledge graph from your conversations without calling an LLM, connecting decisions into causal chains to understand not just what you chose but why. The soul distillation feature accumulates these drifts and builds a living persona rather than a static snapshot. Instead of just knowing that you mentioned something three days ago, the system knows you mentioned it, what decision you made, and how that fits into your longer pattern of behavior. It distinguishes between what you said, what you decided, and how your decisions have shifted across time. Search works across four dimensions at once: full-text search, an inverted index, TF-IDF scoring, and a concept called Shadow Sand (a secondary index used to weight results based on your context and history). The system also handles Chinese and English automatically, detecting the language of input without configuration. The README is bilingual, with the Chinese sections going into more detail about the architecture than the English sections. Performance benchmarks show write operations completing in about 2ms and searches in 1 to 3ms on a dataset of 3,549 memory entries running on a mid-range laptop. The project can start with a single Docker Compose command or run as a local Python process.