Memoryful guardrail · Intention-time detection
AI agent security has an amnesia problem.
We cure it.
Threat memory on intent and the model's own computation, so AI threats can't hide between messages.
The three surfaces of an agent
SIEMs cover events.
We watch intent and computation,
and turn them into memory.
A guardrail reads one message, decides, then forgets. SIEMs see everything, but only after the agent has already acted. The two surfaces in between are where attacks actually live.
What is steered.
We compress every prompt stream into a small, bounded memory of intent. Cross-session attacks live in the aggregate — we see them before the agent acts.
How it computes.
The model already produces a fingerprint per request. We recycle it into computational memory. "Is this harmful?" becomes "does this match normal?"
What it did.
Tool calls, file writes, logs, network flows. Mature and well-covered by SIEM, UEBA, and agentic SIEMs — but only after the agent has already acted.
Memory persists across sessions · across agents