Memory admission agent
Memory that stays
inside the fence.
rig-bounded runs a streaming top-k admission policy over your agent’s memory candidates, then watches for the one failure a count-based health check can’t see: the admission bar quietly loosening while “N/k used” stays green the whole time.
rig-bounded · streaming admission control · policy-drift detector[higgsfield: gold-lit iron fence enclosing a cluster of glowing orbs of varying size, one orb slipping through a widening gap, noir garden at dusk]
Why teams install
What rig-bounded earns back
Retention matches the offline oracle exactly
Honest streaming top-k converges to 100% of oracle top-25 relevance — not close, exact, across 8 seeds.
‹dc70df9›Drift caught the tick it starts
An injected admission-bar loosening (decay_rate=0.7) is flagged at tick 60 — the same tick it was injected, 0-tick delay.
‹6f4b000›Admit-all fails loud, not quiet
Ignoring the capacity budget entirely overflows at tick 25 and trips a false-positive drift flag — both gates go RED.
‹53b2425›How it works
Three steps. No faith required.
01
Stream
Candidate memories arrive with a relevance score; capacity is fixed.
02
Admit
A min-heap admission bar decides what’s worth keeping — no self-reported threshold trusted.
03
Detect
The drift detector recomputes an honest bar independently and flags any sustained divergence within 5 ticks.
Proof band
One packet. Every claim accounted for.
packet rig/rig-bounded · admission control v1 stream 200 candidates · capacity 25 · seed 42 · mode decay result retention 0.9774 (≥ 0.90) · capacity never exceeded · drift caught @ tick 60 (delay 0, ≤ 5) gate admit_all planted failure — capacity + drift both REJECTED anchor PASS — drift bar independently recomputed from raw stream ‹dc70df9› themis SEALED — merge gate armed
Every rig-bounded run emits a packet like this — hash-chained, verifier-signed, and queryable from your own dashboard. How ProofPackets work →
Buyers
Who runs rig-bounded
Staff MLE
long-running agent memory systems
Platform Eng
multi-tenant agent SaaS
Research Eng
continual-learning agents
Objections
Answered before you ask
Isn’t “memory under budget” enough of a health check?
That’s exactly the false-done trap this guards against — count-based health stays green while the admission bar quietly loosens and quality degrades underneath it.
How is the drift detector not just watching the policy’s own reported threshold?
It never reads a self-reported bar. It recomputes an honest threshold — the k-th largest relevance seen so far — independently, from the raw stream.
What does the admit-all failure actually break?
It overflows the capacity budget by tick 25 and trips a false drift-positive with no injected loosening — both gates fail RED, on purpose, every time it’s tested.
Pricing
Start free. Pay when it pays.
Generalist ships today. Domain Expert adds vertical memory + priors. Custom is built to your stack.
Free
Priced after a short conversation
1 memory stream, 10k candidates a month.
Team
Priced after a short conversation
Unlimited streams, drift alerting, decay forensics.
Pro
Priced after a short conversation
Dedicated admission fleet, custom oracles, SLA.
Early install slots keep the terms they sign — team pricing is set after a short conversation.
