Drop-in memory infrastructure that retrieves the right fact, routes every query to the right mode, and grounds every answer in provenance.
Agents pull a verified fact in ~100ms instead of re-deriving it over ~30 seconds. The redundant work you stop doing is the line item that moves this quarter.
A frozen sub-50ms classifier routes each query to one of five cognitive modes — trivial lookups go fast and cheap, high-stakes recall goes deep.
Every recall returns with confidence, sources, and a tamper-evident hash. A record of what was known, and when — audited, not asserted.
Behind the firewall the waste is worse than on the open web: one rep's agent researches a customer Monday, another's repeats it Thursday, legal re-derives the same clause. The same fact, re-derived and re-billed, team after team.
Forcing every question through one retrieval strategy is like using one tool for every job in the workshop. We read each query and route it to the strategy it actually needs.
Accuracy, latency, cost — push toward one corner and you give ground on another. So we don't pick one point and live there; the classifier picks a different point for every query.
LoCoMo and LongMemEval are saturated — every serious system clusters at the top. MemoryAgentBench is adversarial and multi-session; on multi-hop conflict resolution, most systems collapse below 7%.
How we report. Backbones vary across systems; our 62.19 exceeds the highest overall result in the paper (49.6). LoCoMo and LongMemEval are saturated, so we report our own measured figures rather than disputed self-reported numbers.
A frozen sub-50ms classifier reads intent and picks the cognitive mode.
Four retrievers plus a version-aware graph select the candidates.
MemoryRank weights by reliability and recency; stale facts decay.
A synthesis pass at temperature zero composes the answer.
Confidence and provenance return; the hash anchors to the audit trail.
No system can declare a fact true — so we don't. What gets anchored is compact and tamper-evident: the hash, the timestamp, and the trail of how a claim's confidence evolved. Run it as a private audit chain, or anchor to Supra L1.
Same SDK and API across every deployment — start managed, move private later, rewrite nothing.
Fully managed, autoscaling memory service. The fastest path to production — usage-based, 99.9% SLA, zero infrastructure to run.
Deploy into your VPC, on-prem, or fully air-gapped. BYOK encryption and a private provenance anchor; data never leaves your environment.
Python and Node SDKs, a REST API, and framework plugins. Add verifiable memory to an existing stack without rebuilding your pipeline.
| Capability | Managed cloud | Private | API & SDK |
|---|---|---|---|
| Classifier-routed retrieval | ✓ | ✓ | ✓ |
| Hash-chained provenance | ✓ | private anchor | ✓ |
| Autoscaling & 99.9% SLA | ✓ | your infra | — |
| BYOK · data residency | ✓ | ✓ | — |
| Air-gapped option | — | ✓ | — |
| SSO/SAML · dedicated support | enterprise | enterprise | — |