Engram packs local LLM knowledge graph

This is less a chatbot than a local reasoning substrate, and that is the right direction if you want LLMs to accumulate durable knowledge instead of churning answers. The catch is obvious: the system looks only as good as the models you plug in, so the “single binary” story is strongest when the inference quality holds up.

• –The fact-extraction loop is the most credible part: deterministic NER up front, then LLMs only where structure and judgment are needed
• –Confidence decay plus contradiction correction is a pragmatic way to keep a graph from fossilizing into stale assertions
• –The debate modes are compelling for analysts, but they will be very sensitive to model size, context window, and prompt discipline
• –MCP support matters more than usual here because it turns the graph into an agent-accessible memory layer, not just a standalone app
• –The local-first packaging is the differentiator: no DB server, no Docker stack, just a portable knowledge store you can back up by copying one file