Ask a generic AI assistant about your project and it answers from its training data — confident, fluent, and disconnected from anything you've actually written. The Lounge agent works differently: it answers from your notes, and the interesting engineering is in how it decides which notes matter.
Search finds the entry points
When you ask a question, the agent first runs retrieval over your workspace — full-text matching plus semantic similarity — to find candidate notes. This is the part most 'chat with your notes' tools stop at, and it's why they fall apart on questions whose answers are spread across several documents.
Retrieval gives you the notes that mention the question. It misses the notes that answer it.
Links provide the context
This is where your wiki-links earn their keep. From each entry-point note, the agent walks the link graph: notes it references, notes that reference it, and the tags they share. A question about a product decision might start in a spec, hop through [[meeting notes]] where the trade-offs were argued, and land in a customer-call summary that explains the why.
The traversal is bounded and ranked — links from recent, frequently-connected notes weigh more than stale ones — so the agent assembles a compact working set instead of dumping your whole vault into context. The graph you built by hand acts as a relevance signal no embedding model can fake, because it encodes what you considered related.
Grounded answers, visible sources
The agent composes its answer strictly from that working set, and every response cites the notes it drew from. If the answer isn't in your notes, it says so rather than improvising. An assistant that's occasionally wrong but always checkable beats one that's usually right but opaque.
The upshot: the agent is exactly as smart as your graph is connected. Every link you add while writing is also a path the agent can follow later. Thinking in connections pays you twice.