health-yang/mineecho

MineEcho is a local-first framework for building a personal AI assistant that actually remembers things across conversations. Unlike a regular chat interface that treats each conversation as fresh, MineEcho keeps a structured memory of past interactions at multiple levels: raw chunks from recent sessions, daily summaries, weekly summaries, and monthly archives. When you ask something, the system pulls relevant memories by combining keyword matching, semantic similarity, and recency scoring before deciding what context to include. The knowledge base works similarly. When you import documents, notes, or other sources, MineEcho turns them into browsable wiki pages, chunk indexes, and graph nodes rather than hiding everything in an opaque vector store. At query time it searches across four channels: semantic vector search, keyword matching for both English and Chinese, structured fields like titles and tags, and a graph channel that can follow one-hop relationships between entities. This makes it easier to inspect what the system knows and how it found something. Skills and external AI apps are handled through a shared routing surface. You can register a local function, import a packaged skill, or connect an external AI application, and they all end up in the same registry. The system derives routing triggers from names and descriptions and scores candidates before returning the best match. This means external apps behave like local capabilities rather than becoming separate silos you have to manage individually. A built-in compression layer called TokenLess reduces the amount of context that gets passed to the language model. It applies scenario-specific rules for things like build tool output, package manager logs, and document extraction results, keeping errors, counts, and actionable lines while dropping repeated or low-value content. The system records how much reduction actually happened locally so you can see real numbers for your own workload. Setup requires Node.js 22 or later. You clone the repository, run a single install command, copy an environment file, and start a development server. The console runs at a local URL where you configure your model provider keys. Runtime data, chat history, and keys all stay on your machine and are excluded from any release export. The project is source-available and includes documentation in both English and Chinese.