Mem

Memory systems, readers, and writers.

Study method

What each comparison tests, what stays fixed, and how to read its measures.

Change studyLuna reasoning effort

Study

13 reader + 4 writer inputs · 53 calls including calibration

Reader · 13 inputs: Thirteen identical saved questions, prompts, and retrieved contexts. Completeness and exact quote binding remain separate checks.
Writer · 4 inputs: Four identical extraction inputs, including English and Chinese conditional decisions. All profiles preserve the two tested conditional-scope and owner samples. Subject and effective-time omissions remain.

Reader completeness and quote binding are separate checks. Writer resource use does not establish a semantic quality ranking.

Inputs, shared settings, and execution environment

Controlled synthetic text fixtures with real provider calls. Saved outputs are historical observations, not simulated scores.
GPT-6 Luna, OpenAI provider, no fallback. Effort is the only explicit model parameter changed; temperature is omitted.
Full 128,000 output allowance for every call; every response ended normally. Thirteen identical reader inputs and four extraction inputs per profile, plus two separate calibrations.
Three blocks rotate profile order. Actual cache-read and cache-write tokens remain in the original metrics. Reader contexts use the retained Qwen3-Embedding-8B 4096-dimensional retrieval.
Fifty-three calls; new settled spend $0.016351425. Reader and writer costs in the tables exclude the two calibrations.

What this study can and cannot support

A small diagnostic pass, not a general intelligence ranking or a statistically stable quality order. Repeated questions share facts and are not independent projects.
A matching literal quote proves source binding. It does not prove semantic entailment, complete answers, tenant isolation, or customer acceptance.
Missing or infrastructure-invalid measurements remain unavailable. They are not semantic zeros. Research evidence is neither Calibration nor Official Core evidence.
No native storage, reconciliation, retrieval, or final Agent tool loop was rerun.
None saved about $0.04 per thousand reader answers in this workload. Four writer samples do not justify a routing policy. No default configuration changed.
The original report suggested exploring none for routine extraction. The later interpretation shown here retains one medium default. The original report remains byte-identical.

3 profiles · recorded model, version, and settings

Exact configurations in the selected study
Profile	Version or model	Configuration
Luna low	`openai/gpt-6-luna`	Settings provider OpenAI; no fallback effort low outputAllowance 128,000 tokens
Luna medium	`openai/gpt-6-luna`	Settings provider OpenAI; no fallback effort medium outputAllowance 128,000 tokens
Luna none	`openai/gpt-6-luna`	Settings provider OpenAI; no fallback effort none outputAllowance 128,000 tokens