AI agents are throwing away their own experience after every run — and a new database architecture proposal wants to fix that.
Researchers have published a paper introducing Trellis, a data layer that treats what they call an "experience graph" as structured, queryable state rather than disposable session logs. When an AI agent spends hundreds of steps generating code, running tools, hitting failures, and branching into alternatives, that entire search history — artifacts, outputs, failure signals, and causal links — is currently flushed or buried in JSON checkpoints that can't be queried across users or recovered after a crash. Trellis proposes storing all of it in a governed database, turning operations like finding a promising prior agent run into a graph retrieval query, and extracting training data into a materialized view.
The practical upside isn't abstract. At Meta, a production kernel optimizer called KernelEvolve that uses cross-session reuse reportedly hits target speedups roughly 10 times faster while cutting token costs by 52%. Those numbers suggest that agents repeating work already done by other agents — or by themselves — is a real and measurable drain, not a theoretical inefficiency.
The analogy the authors reach for is telling: logs made databases reliable; experience graphs may make agents cumulative rather than amnesiac. That's a reasonable framing, though it's worth noting that the paper comes from arXiv and the production evidence is a single internal Meta system — so treat the 52% figure as a promising data point, not an industry benchmark.