The Almanac launch pitches an agent that “just knows you,” but strip the demo away and it’s a bet on one specific RAG design choice: pre-compile the knowledge into wikis instead of retrieving it at query time. That’s a real architectural axis — fast and coherent, but stale-prone — not marketing.
- 🎯 The claimed moat is memory; the moat that matters is the permission boundary between the personal wiki and the shared company one — one leak across it and enterprise adoption stops cold.
- ⚠️ Pre-compiling doesn’t kill the RAG problem, it relocates it — freshness now lives in the wiki-refresh pipeline instead of the retriever, and stale context fails silently.
- 🔍 Provenance back to source documents is the right call; a compiled wiki with no citations is unauditable the moment someone asks where an answer came from.
- ⚡ Long-horizon, session-spanning tasks are the genuinely hard part — most agents are session-bound, and holding project state across days without drift is still unsolved.
- 📊 Connector sprawl (Gmail, Calendar, PostHog, and friends) is where these systems rot operationally, long before model quality is the bottleneck.
The HN discussion is already circling privacy and permission depth, which tells you where the real evaluation lives. My bet: the winner in company-brain agents won’t have the best model or the biggest wiki — it’ll have permission-aware retrieval a security team will actually approve.