AI Safety · Part 3 of 7

Grounding: memory beats guessing

Watch: what grounding measurably does · 61 seconds · animated, sound on

Measured, not promised

When Stanford researchers tested models on legal questions, grounding them in real document databases cut hallucinations by more than half. Not to zero. More than half is worth building around; zero is a lie.

Two very different machines

Ask an ungrounded model about your business and it fills the silence from patterns. Ask a grounded one and it reads your actual documents and memory before it answers. Same question, completely different failure modes.

The measurement

Stanford researchers asked models legal questions at scale. Without grounding, they hallucinated on 58 to 88 percent of them, depending on the model. Tools grounded in real, curated document databases cut that to 17 to 33 percent. That is the honest shape of the win: grounding cut errors by more than half, and the best grounded tool still got one in six wrong.

What this means for your agents

Memory is grounding applied to your own work. Without it, a model asked about last month's decision has two options: admit it does not know, or invent your history. Cronus agents share one memory system, Hindsight, and it changes the math:

The honest limit

Memory prevents the "inventing your history" class of hallucination. It does not make a model incapable of being wrong about the outside world, and a memory system will faithfully store a mistake if a mistake gets written down. That is why memory you can see matters as much as memory that exists.

Sources: Dahl et al., Large Legal Fictions, 2024 · Stanford RegLab, Hallucination-Free?, 2024.

← Part 2: Context Part 4: Actions →