Most memory tools for LLMs just append to a flat list. Ember uses a different approach: a Shadow-Decay system that detects when new information contradicts stored memories and marks old ones stale automatically, rather than piling up conflicting facts.

It runs locally, stores everything in a vector database you control, and plugs into Claude (and other MCP-compatible clients) through 16 tools covering store, recall, graph search, and health monitoring.

The core problem it solves: you tell your AI you changed jobs, and three weeks later it still references the old one. Ember's contradiction engine catches that at write time.

Repo: https://github.com/Arkya-AI/ember-mcp

Still alpha — single user, some token overhead, scoring weights need tuning for different use cases. Posting here for technical feedback before wider release.