Enterprise AI agents are stalling, but not because of model performance. The bottleneck is permissions. Every agentic workflow hits the same wall: what is this agent allowed to touch, on whose behalf, and how does the system know? Workday's answer is Sana, a system that turns its HR and financial record-keeping into a governance layer for agents. As explained to VentureBeat, Sana ensures the integrity of approvals and security models, solving the chaos of DIY AI approaches.
Accuracy matters most
In HR and finance, almost right is not acceptable. Paying people correctly, closing the books, or managing work schedules demands absolute fidelity. Sana uses Gemini as a conversational surface but authenticates and authorizes through Workday's identity model. Audit trails stay within the customer; Gemini only retains interaction logs. Governance must live inside the system of record, not outside.
Sponsored Protocol
A new memory for LLMs
Meanwhile, the MeMo framework from MIT and other universities offers a different path. Instead of retraining expensive LLMs, MeMo encodes new knowledge into a separate, smaller MEMORY model that talks to a frozen EXECUTIVE model. Instead of RAG or fine-tuning, it uses reflections — question-answer pairs covering every angle of a corpus. Tests show up to a 26% performance jump just by swapping the executive model, without retraining memory.
What it means for enterprises
MeMo's real value is continuous updates without catastrophic forgetting. Using model merging, companies can introduce new policies without full retraining. Limitations exist: initial memory training is expensive (roughly 240 GPU-hours on H200s for data generation) and provenance is less traceable than traditional RAG. But for synthesis across complex documents, MeMo dominated benchmarks like NarrativeQA with 53.58% accuracy versus HippoRAG2's 23.21%.
Sponsored Protocol
As AI agents must handle sensitive data, the combination of structured governance like Sana and modular memory like MeMo could define the next standard. Enterprises will have to choose between governance and synthesis, but the direction is clear: enterprise AI is not just about more powerful models.
