Measured Systems Results

Memory should not scale like the problem.

ExergyNet measures the work required to retrieve, reason over, and act on persistent AI memory, then redesigns the control plane to reduce unnecessary computation without surrendering evidence integrity.

Explore the Results White Paper - Coming Soon
11.3×
Correct-task throughput
Measured H200 full-context comparison.
52,753.8×
Median paired retrieval speedup
Tested 1GB / 100-query enterprise population.
1.886×
Adaptive router vs always-legacy
Frozen LNES-86.2 holdout population.
C16
Concurrency validated
Isolated VMN routing and evidence checks.
Benchmark Families

Separate experiments. Separate envelopes.

These results are not combined into one synthetic score. Each card states the measured result, the principle it supports, the test envelope, and the known limitation.

xLMP / Accelerator Efficiency

11.3× more correct-task throughput than full-context replay at each method's highest tested SLA-compliant operating point on the documented H200 workload. The same evidence package reports a bounded 660-820 active-token xLMP prompt range as stored memory grew past the model context window.

EnvelopeSingle NVIDIA H200, Nemotron model, synthetic memory corpus. MeaningLess irrelevant memory entered active model context. LimitProxy wall-time result, not integrated energy/GPU-seconds.

Compact Deterministic Index

87% to 94% symmetric correctness and 52,753.8× median paired speedup in the tested 1GB / 100-query enterprise population. The indexed path was faster on all 100 paired queries in that specific run, with 3.96× minimum observed speedup.

Memory work compression funnel Stored memory is reduced to posting candidates, scored candidates, and evidence. MEMORY SELECT SCORE EVIDENCE C P(q) S(q) E(q)
Envelope1GB corpus, 100 enterprise queries. MeaningEvidence selection reduced retrieval work in this population. LimitCandidate selectivity is workload-dependent.

Adaptive Deterministic Work Routing

No single retrieval strategy won on every workload. LNES-86.2 selected among retrieval paths before expensive work and measured 1.886× vs always-legacy, 1.142× vs always-compact, 1.150× vs the first-generation router, 95.6% holdout route-selection accuracy, and 1.023× oracle gap.

Adaptive deterministic work routing A query flows through work analysis into either direct or indexed retrieval before verified evidence. QUERY WORK ANALYSIS DIRECT INDEXED
EnvelopeFrozen 231-query routing population; 45 holdout cells. MeaningWorkload-aware routing approached oracle-best selection. LimitProduction remains blocked pending concurrency-tail hardening.

What We Learned

Indexed retrieval is not universally faster. VMN-native validation showed aggregate median speedups below 1.0× at both 10MB and 100MB, while larger/filler-heavy subsets improved. That negative result led directly to adaptive deterministic routing.

10MB0.72× aggregate; 0.69× small/entity subset; 2.28× filler/large subset. 100MB0.70× aggregate; 0.68× entity subset; 2.11× filler/large subset. LimitTwo scale points; not a complete crossover surface.
Charts

Make the evidence inspectable.

The diagrams below are sized for reading, not decoration. Click any chart to expand it for closer inspection.

Click to expand LNES-84.3 Phase 8-17 VMN validation chart sheet

Imported from the sealed Phase 8-17 chart artifact. The sheet shows the workload-dependent VMN finding, concurrency behavior, mutation/restart/corruption checks, and the production-promotion gate.

Retrieval Speedup Envelope

Compact Index retrieval produced very large speedups in the 1GB enterprise population, while VMN-native validation showed workload-dependent slowdowns in other populations.

Click to expand Retrieval speedup envelope Bars compare 1GB enterprise median speedup with VMN aggregate and filler subset results. Measured Retrieval Speedup Envelope Log-compressed visual scale; labels show exact reported figures. 1.0x parity 1GB enterprise 52,753.8x VMN filler 10MB 2.28x VMN filler 100MB 2.11x VMN aggregate 0.70-0.72x

Adaptive Router Lift

LNES-86.2 improved over both fixed baselines and stayed close to oracle-best routing in the measured holdout envelope.

Click to expand Adaptive router lift Bars show speedup of the adaptive router against baselines and its small oracle gap. LNES-86.2 Adaptive Router Lift Frozen holdout result; production gate remains blocked pending concurrency-tail hardening. 1.0x parity vs legacy 1.886x vs compact 1.142x vs router v1 1.150x oracle gap 1.023x
Resilience and Authority

A memory system must stay correct while state changes.

PASSMutation consistency
PASSRestart persistence
PASSCorruption recovery
PASSDeterministic rebuild
PASSAuthority separation
C16Concurrency validated in isolated testing

Right to know is not right to act.

Retrieved memory can inform a decision. It cannot manufacture permission to execute one. Memory control, model reasoning, and execution authority remain separate planes.

Memory and authority boundary Memory/evidence produces knowledge; authority/consequence controls action. MEMORY / EVIDENCE Knowledge RIGHT TO KNOW != RIGHT TO ACT AUTHORITY Action
Terminology

A new system needs precise language.

xLMPThe AI Memory Control Plane.
Work CompressionReducing persistent memory into the minimum evidence work required for a task.
Correct-Task ThroughputRate of successfully completed tasks, not raw token generation.
Evidence ParityWhether alternative retrieval paths return equivalent authoritative evidence.
Adaptive Deterministic Work RoutingSelecting among retrieval mechanisms using pre-execution workload state.
Oracle GapDistance between measured routing cost and an oracle that already knows the faster path.
Authority SeparationMemory can inform an action without granting permission to perform it.
Persistent MemoryDurable memory that survives inference sessions and process boundaries.

The white paper is coming. The evidence is already visible.

These public results publish the phenomenon. The forthcoming paper explains the architecture without exposing the private operating recipe.

White Paper - Coming Soon Claim Ledger