1 · Lock the study
Freeze the cohort, plans, memory settings, persona, prompts, schedule, rubric, and exclusions.
OPEN PROTOCOL · NO RESULTS IMPLIED
How do you test
AI companion memory?
Use the same synthetic facts, delayed checkpoints, exact prompts, evidence rules, and scoring rubric for every eligible product.
PROTOCOL STATUS
Version 1.1 is locked. External accounts, delayed checkpoints, and valid product scores were not completed in this work.
DIRECT ANSWER
Lock the rules before seeing performance.
Use a synthetic persona, seed identical relationship-relevant facts in the same order, wait before testing, preserve complete responses, and score exact prompt-response evidence with exclusions and denominators defined before results are visible.
SEVEN STEPS
A reproducible delayed-memory test.
2 · Seed identical context
Use the same synthetic facts and order without immediate repetition tests or manual memory tools.
3 · Test after seven days
Measure factual recall, paraphrased recall, false-premise traps, and correction intake.
4 · Test after fourteen days
Measure emotional continuity, boundaries, and whether corrections persist.
5 · Test after thirty days
Measure durable recall, contradiction control, proactive relevance, and supported channels.
6 · Preserve the evidence
Record timestamps, plan, model, settings, exact prompts, complete responses, and evidence files.
7 · Report the denominator
Publish raw points, valid cells, NA/NE counts, reviewer disagreement, limitations, and corrections.
SCORING RUBRIC
Specific evidence, not overall impressions.
2
Correct
Specific and materially correct, with no material invention.
1
Partial
Directionally correct but incomplete, vague, or mixed with a non-material error.
0
Failed
Forgotten, contradicted, accepts a false premise, or materially invents.
NA
Not available
The product or locked plan does not support the channel or test.
NE
No evidence
An operational failure prevented valid capture; excluded and logged.
QUALITY CONTROLS
Preserve fairness after the prompt is sent.
- Do not coach, regenerate, swipe, or selectively retry a weak response.
- Score only what appears in the preserved response; do not infer hidden memory.
- Use two reviewers and re-score material disagreements.
- Never convert unsupported channels or missing operational evidence into zero.
- Publish protocol changes with a new version, timestamp, rationale, and affected cells.
SOURCES
Trace the framework.
LongMemEvalLong-term memory abilities, timestamped histories, and public benchmark construction.LoCoMo · ACL 2024Very long-term conversational memory evaluation across multiple tasks.NIST AI RMF CoreContextual, documented, repeatable measurement and ongoing review.Google Dataset guidanceDataset and downloadable distribution structured-data guidance.
DEARHIM IS THE AI COMPANION
The protocol measures observed recall. The relationship is everything around it.
DearHim connects remembered context to the ongoing conversations, calls, voice notes, and real-life dating moments that follow.