Show HN: Ember MCP – local persistent memory for LLMs, kills stale memories (github.com)

Ember MCP has launched as a local-first memory server for large language models (LLMs), employing Voronoi partitioning to manage context and minimize hallucinations from stale data. This innovation allows AI to maintain a permanent memory that seamlessly carries over across different interfaces, enabling users to shift from environments like Claude to Cursor without losing continuity. Most notably, Ember automatically discards outdated information, ensuring that AI-generated code and recommendations reflect the current technology stack, effectively preventing the common issue where AI suggests obsolete solutions. The significance of Ember lies in its ability to provide long-term memory without reliance on cloud infrastructure, maintaining 100% privacy for user data by running locally. Key technical features include drift detection, which automatically flags outdated memories and uses statistical monitoring to manage the relevance of knowledge in real time. By clustering knowledge and tracking data freshness, Ember addresses the frequent problem of “semantic collision” in evolving projects, enhancing the fidelity of AI responses. With tools for capturing, recalling, and contextualizing information, Ember equips developers with a robust mechanism to engage with their AI, fostering a more dynamic and accurate interaction experience.