Memory Doesn't Equal Expertise: Builders Ask Where Domain-Specific Agents Get Their Knowledge
sartomiki · reddit · 2026-09-10
The poster observes an awkward plateau in agent tooling: persistent context is starting to work — memory layers, vector stores, and MCP servers holding state across sessions — but few agents show real depth in a single domain. Most remain a generalist model with better memory.
His core argument: memory eliminates re-explaining, not the lack of expertise. An agent that remembers every contract you pasted still isn't a legal expert; it guesses once you ask something beyond what you provided.
He asks the community: where does knowledge for domain-specific agents actually come from (your own docs, curated datasets, fine-tuning)? Where does it fall short? And is "generalist + good retrieval" honestly enough for most use cases?
More from coding & agent
- Codex tip: don't let Astra manage other threads — usage dropped from ~70%/day to 25% — jdjohnson · 2026-09-10
- Resend integrates with Vercel v0: add email to your project in a few clicks — tomjohndesign · 2026-09-10
- Interactive blog maps the fragmented 3D ecosystem: why 3D has no single format — HankYeomans · 2026-09-10
- Three takeaways from the Navier-Stokes situation: agents aren't there yet — pranavmarla · 2026-09-10
- GitHub adds enterprise-managed permissions for Copilot agent ops: shell, files, network — PaulShellDev · 2026-09-10
- Site built in 12 hours on a Mac Mini by orchestrating multiple AI agents in parallel — iannuttall · 2026-09-10