Veridical Tech — a memory systems lab

Beyond the limits of AI memory.

We build and operate the SPMOS external-memory system and its products, run experiments that give AI bots durable memory and identity, and work toward a general memory kernel for world and physical models.

Explore SPMOS.ai ↗Three fulcrums ↓

SPMOS is a governed long-term memory system for AI. Our ongoing mission is to enable AI to deliver value its users can count on — continuously, reliably, and safely.

A night desert under the Milky Way — a holographic star map projected from a stone, charting the route from a shipping product toward research

SPM-Polaris

The current production SPMOS product — for individuals and enterprises alike.

Governed long-term memory for AI agents. Connect through the Hosted Provider Proxy, Local Proxy, or MCP for evidence-gated recall and deterministic compression. Fast recall adds no generative LLM call; deep evidence collection runs only when needed — and shows up on the receipt. spmos.ai ↗

Aleph — beginning · Shipping · spmos.ai

A sunlit desert base camp where AI bots build and work together — a dog operator injects long-memory modules through a holographic interface

Memory & identity for AI bots

Durable task state and a stable identity boundary let an AI bot carry long-horizon work across sessions and tools — without an ever-growing transcript.

Mem — continuation · Experiments running

A lone robot crossing a golden desert toward a giant planet marked with the triangle and Hebrew letters — the terminus of the arc

A memory kernel for world models

SPM-EVA — the SPMOS engineering prototype of a general long-term memory kernel for world models.

Persistent state, evidence, time, and controlled revision beneath any model class that has to act in a changing physical world.

Tav — terminus · Open research problem

SPM-EVA · Engineering prototype

An external, machine-native multimodal memory kernel for world models. The world model perceives; SPM-EVA owns canonical identity, lineage, and recall.

Simulation

Synthetic reference-VM retrieval: 10,000,000 keys; 0 candidate false negatives · integrated unique-recall worst p99 2.471 ms
Full-duplex development: 7,670 occurrences; Hot p99 0.681 ms · Long p99 0.421 ms · 12/12 gates passed

Public benchmark — EPIC-KITCHENS-100 validation profile

Public audiovisual core replay: 8 videos / 32 occurrences; 32/32 deterministic opaque bodies · byte-identical rebuild
Public validation: 138 videos / 38,350 queries; All hard correctness counters zero · 851/851 ambiguous queries refined to a unique match
Public validation latency: same profile; Long-unique p99 54.630 ms — above the frozen 20 ms gate. A valid negative, disclosed.

Function II — record authority

Memory hides two functions — compute substrate and record authority. The field optimizes the first and assumes the second; SPM-EVA occupies the second.

Endogenous memory has three blind spots no scale can fix: it dies with the model, it cannot prove which version said what, it cannot travel across organizations.

Failure here is explicit, not silent — ambiguous recall returns REFINE_REQUIRED, never a confident wrong answer. The falsification conditions that would kill this approach are written down in advance.

A mechanically validated engineering prototype — not a finished product. The failing latency tail has been localized, not yet repaired. Detailed reports are shared under controlled technical review.

Companies and research institutions working on world models — we welcome joint testing. Working with the lab · contact@veridicaltech.com

We read everything. We answer what we can.