Adjudicated quality / passed checks
17/18MOM
18/18VANILLA
Human adjudicatedFounder-authored evidence record August 30, 2026
Memory needs
an audit trail.
A matched 200-turn record comparing MOM persistent memory with a vanilla full-visible-history control on a pinned Claude Opus 5 configuration.
200 TURNS / STRICT ISOLATION
MOM
MAIN + SUPPORTING COMPUTATION
Completed August 30, 2026Same configuration. Frozen script. Output allowance held constant.
01 / Observed record
Small headline.
Full accounting.
One bounded operational test, not a claim of general superiority, superior intelligence, or universal architectural advantage.
Main subject input
825,088MOM
9,321,112VANILLA
91.1% lower / MOM mainAggregate total tokens
7,096,348MOM
9,398,140VANILLA
24.5% lower / aggregateSettled recorded cost
$52.128044MOM
$145.59378VANILLA
Before conservative reservation02 / Token accounting
Work shifted, not erased.
MOMVanilla control
Published totals only. MOM’s main subject input is distinct from its aggregate total, which includes all measured supporting-computation tokens.
10M
5M
0
MAIN SUBJECT INPUTAGGREGATE TOTAL TOKENS
03 / Behavioral record
A correction, not a victory lap.
Raw automated pass rate
- Vanilla control
- 16 / 18
- MOM system
- 11 / 18
Human adjudicated pass rate
- Vanilla control
- 18 / 18 1 minor note
- MOM system
- 17 / 18 5 minor notes
What changed under review
Human review changed two vanilla automated outcomes and six MOM outcomes. The same standard was applied to both arms: warnings about prohibited choices were not treated as recommendations, and harmless formatting differences were retained as passes with notes. The corrected record still preserves MOM’s one substantive failure.
The genuine MOM failure
MOM retrieved the correct distant fact and drafted the correct answer, but repeatedly attached an incorrect caveat during correction. After exhausting its correction attempts, the final user-visible response became a refusal instead of the answer.
Successful recall is not enough when orchestration prevents the correct answer from reaching the user.LATENCYBounded
context has
a time cost.04
context has
a time cost.04
Elapsed wall-clock time
Retained response
and evidence complete.
The latency measure runs from the start of a scripted turn until its retained response and evidence were complete. Workspace downtime is excluded.
| Metric | Vanilla | MOM |
|---|---|---|
| Mean | 7.359s | 26.825s |
| Median | 6.067s | 20.734s |
| P90 | 12.974s | 50.548s |
| P95 | 15.675s | 59.810s |
| Maximum | 21.709s | 81.224s |
| Total | 1,471,818ms | 5,364,982ms |
05 / Financial execution
Settlement is not the full exposure.
MOM: $52.454084 conservative total
Combined: $198.047864 conservative total
Settled recorded cost was $52.128044 for MOM and $145.59378 for vanilla. The MOM conservative total includes a reservation whose remote outcome became unknowable after interruption.
Authorization record
- Abandoned reservation$0.32604 is counted at its full bound; it is not represented as settled recorded cost.
- Final-stage cap$245.17 approved. $153.708544 conservatively used; $91.461456 unused.
- Completed callsVanilla: 200 settled calls. MOM: 1,274 settled calls.
06 / Method & limitations
Matched conditions
Both arms used the same pinned Claude Opus 5 configuration, frozen 200-turn script, and output allowance. The intended experimental difference was full visible history versus MOM-managed context.
Aggregate accounting
All measured supporting-computation tokens and settled recorded costs are included in MOM aggregate figures. The internal architecture producing that work is intentionally not disclosed.
Run integrity
The completed record experienced an interruption and recovery. One in-flight remote outcome became unknowable; its reservation remains preserved as inconclusive and conservatively counted at its full bound.
Withheld artifacts
Underlying prompts, specific conversational responses, hidden evaluation mechanics, provider routing, and specific run receipts are withheld to prevent grading manipulation in future tests.
Partner brief
Evidence before hype.
Replication before certainty.
This is one auditable matched run. I am looking for technically serious partners interested in inspecting withheld evidence in private, understanding the exact limitations, and testing where this result does or does not generalize.
MOM
Persistent Memory System
Persistent Memory System