Pretrained LLMs may fail at company work because they lack organization-specific memory
thebvg · reddit · 2026-07-22
The author argues that pretrained LLMs behave like a cortex without a hippocampus: they have broad general knowledge, but no fast, organization-specific memory for real company procedures.
The core claim is that this is why agents fail in production. The model can improvise, but company work depends on exceptions, tacit knowledge, and procedures that live in conversations and people’s heads rather than in clean documentation.
The post suggests that retrieval alone is not enough, because it only recalls documents instead of consolidating recurring work episodes into approved procedures. A workable system would need cited, versioned, human-approved workflows plus an audit trail and sign-off on sensitive actions.
The author closes with three questions:
- whether missing company procedures is really the main blocker,
- whether the cortex/hippocampus analogy is useful,
- and what made agents trustworthy enough in real workflows.
More from Companies & People
- Legal Ops Room Finds Zero Confidence in Data, Recommends Data-First AI Rollout — YvesMulkers · 2026-07-22
- Codex set up its own mailbox and negotiated with local vendors for a water softener — sebpaquet · 2026-07-22
- Multica says multica-agent is now its top GitHub contributor — jiayuan_jy · 2026-07-22
- Tesla app decompile shows Optimus home controls, BLE pairing and data capture — CyberRobooo · 2026-07-22
- Programming skills are now being used to 100x content creation speed — jackedAJ · 2026-07-22
- Zhipu website traffic jumped to 10.37 million after GLM-5.2 launched — FinanceYF5 · 2026-07-22