Mem

Memory systems, readers, and writers.

3 studies
6 comparisons
189 original files

All memory comparisons

Complete + bound counts complete answers with source-matched quotes. Each row is a separate test.

Memory comparison overview
Comparison	Complete + bound	Resources	Evidence scale
Readers · fixed-context replay → Luna / DeepSeek Flash 1017	DeepSeek Flash 101711/13 Luna medium11/13	$0.004811–$0.0193694.54–5.05 s / request	13 inputs / profile2 profiles · one pass
Readers · fixed-context replay → Luna reasoning effort	Luna low10/13 Luna medium11/13 Luna none11/13	$0.004149–$0.0047273.63–4.95 s / request	13 inputs / profile3 profiles · one pass
Systems · native retrieval → Native memory systems	OpenViking11/13 RAGFlow rc111/13	$0.006922–$0.024261Time: withheld from comparison	13 inputs / profile2 profiles · one pass
Systems · nightly version check → Native memory systems Viewing	RAGFlow nightly11/13	Cost not reported4.74 s / request	13 inputs / profile1 profile · one pass
Writers · extraction replay → Luna / DeepSeek Flash 1017	Quality not measuredDeepSeek Flash 1017 · Luna medium	$0.001142–$0.0029144.57–8.31 s / request	3 inputs / profile2 profiles · one pass
Writers · extraction replay → Luna reasoning effort	Quality not measuredLuna low · Luna medium · Luna none	$0.000709–$0.0011203.52–7.72 s / request	4 inputs / profile3 profiles · one pass

Separate tests · no combined ranking.

说明

Costs are workload totals in USD; resource ranges span profiles within that row. Quote binding does not establish semantic entailment.

Systems · nightly version check

Native memory systems · RAGFlow nightly version check

↑ All comparisons

Refine comparison

Study

Test

13 inputs / profile · 1 profiles · October 20, 2026

Test method

Compare complete, source-bound answers first, then the resources used.

Results & evidence ↓

Complete + bound

Answers · higher is better

RAGFlow nightly11/13

Mean time

Seconds / input · lower is better

RAGFlow nightly4.74 s

Cost not recorded.

Results and evidence

Sort columns. Open Details for the profile and original record.

Expand a row for all metrics

Sort

Metric

October 20 · version check. Values are specific to this test. $0.008500545 is the full new spend for 45 calls, including calibration, fresh embeddings, and the separate Agent probe. It is not a reader-only cost.
Profile version						Evidence	Details
RAGFlow nightlyv1.0.0-rc2-48-g7eb673bfc	11/13	11/13	28/28	4.737 s	—	Result evidence RAGFlow nightly October 20 · version check · 13 inputs per profile Profile Complete + bound answers Cost USD total RAGFlow nightlySelected 11/13 Not measured Complete + bound requires a complete answer and exact quote binding. Other test groups remain separate. All recorded measures Complete answers 11/13 Complete + bound answers 11/13 Quotes matched 28/28 Time s · mean 4.737 s Cost USD total Not measured Configuration & original record Recorded date 2026-10-20 Version v1.0.0-rc2-48-g7eb673bfc Source 7eb673bfc3ab1cc22532fe23a1e9bb85c34e48bd Image sha256:002f50e053e9c5081610c9b0bb9f07a190cb240e1c983cadc9f0e4364dec9153 Reader openai/gpt-6-luna · medium Embedding Qwen3-Embedding-8B · Nebius · 4096 dimensions Platform Native ARM64 Native retrieval + shared reader; separate version check $0.008500545 is the full new spend for 45 calls, including calibration, fresh embeddings, and the separate Agent probe. It is not a reader-only cost. One diagnostic pass with real provider calls. Sort order is descriptive. Source binding alone does not establish semantic support. Open original formal-summary.json ↗

One pass · — Not measured.

说明

USD totals are rounded to 6 decimals. The value tooltip and original record retain the recorded precision. Sorting is descriptive.

$0.008500545 is the full new spend for 45 calls, including calibration, fresh embeddings, and the separate Agent probe. It is not a reader-only cost.

Recorded finding

RAGFlow nightly: 11/13 complete + bound. This is a separate version check; the other systems were not rerun.

Protocol and metric definitions

A separate ARM64 nightly run on the retained inputs. Its three native Agent probes are outside the 13-question result.

Synthetic inputs · real provider calls · one pass · no repeated-trial interval.

Complete + bound: complete answer with every quote matched to its supplied source. Complete: complete answers, before quote binding. Quotes: matched / emitted quotes, with a separate denominator. Source binding alone does not establish semantic entailment.

Full conditions and limitations

Results