← News

ResearchSources: arXiv · 18 Nov 2026

OPAL engine keeps application memory persistent between AI reasoning sessions

A CPU-resident C++17 graph engine preserves record identities, relationships, and accepted numerical state across restarts, with benchmark results reported on Apple Silicon M2 Pro hardware.

In plain words

The OPAL project is a C++17 engine that runs on a computer's CPU and stores an application's records, links, and accepted numbers so they survive between sessions and after a restart. The authors say this lets applications call AI models only when interpretation, planning, or generation is actually needed, instead of keeping reasoning state alive constantly.

What was announced

The arXiv abstract describes OPAL as a CPU-resident C++17 graph engine that separates persistent application memory from probabilistic reasoning. Applications evolve stored state through defined updates, save canonical checkpoints, and recover continuity after restart, with Bounded Local Generator Classes (BLGC) given as the mathematical foundation for this bounded local evolution. The architecture keeps accepted memory available while reasoning is idle, so models are invoked only when tasks require interpretation, planning, or generation.

On Apple Silicon M2 Pro hardware, the authors report a resident-update median of 0.083 microseconds at 100,000, 1,000,000, and 5,000,000 retained Atomic Memory Vertices, covering complete resident-update calls including validation, indexed lookup, and numerical state mutation; initialization and storage-growth costs are reported separately. Reliability tests covered event replay, conflict rejection, concurrent calls, allocation-failure handling, and fresh-process recovery, and engine-version comparisons produced byte-identical checkpoints. The installed PYAMV Python package passed 17 SDK contract checks and 20 workload checks, while KINDRED's Model Context Protocol interface passed 22 focused checks covering memory operations, authorization, and transport.

What it does not show

  • The evidence is an arXiv abstract only; the full paper's methods, datasets, and peer review status are not available in the supplied passages.
  • All performance figures and test results are self-reported by the paper's authors; no independent confirmation is provided.
  • Initialization and storage-growth costs are excluded from the reported 0.083-microsecond median and are said to be reported separately, so total end-to-end cost is not established here.

For security and IT teams

Scope
The reported benchmarks apply to the current engine on Apple Silicon M2 Pro hardware with a frozen binary and 1,024 prepared targets, not to other platforms or configurations.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.