Mem
Memory systems, readers, and writers.
- 3 studies
- 6 comparisons
- 189 original files
Study method
What each comparison tests, what stays fixed, and how to read its measures.
Change studyNative memory systems
13 questions · October 15–20, 2026
- October 15 · native retrieval
- Same six documents, eight conversation records, two projects, and 13 questions. Each native system supplies context to the same Luna reader.
- October 20 · version check
- A separate ARM64 nightly run on the retained inputs. Its three native Agent probes are outside the 13-question result.
Answer completeness and exact quote binding are checked separately.
Conditions
Inputs, shared settings, and execution environment
Conditions
Inputs, shared settings, and execution environment- Controlled synthetic text fixtures with real provider calls. Saved outputs are historical observations, not simulated scores.
- Six text documents, eight conversation records, two projects, and 13 answer attempts; two before updates and eleven after updates and recovery.
- GPT-6 Luna, OpenAI provider, medium effort, full 128,000 output allowance. Qwen3-Embedding-8B, Nebius, full 4096 dimensions.
- At most six native knowledge hits and six memory hits, with a common 32 KiB context allowance. Actual granularity, source hydration, platforms, and elapsed conditions differ.
Limits
What this study can and cannot support
Limits
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 system speed ranking: execution platforms, recovery, host load, vector reuse, and context sizes differ.
- RAGFlow nightly: three separate native Agent sessions used Memory Retrieval, but all structured references were null; one historical answer mixed later policy. They are outside the 13-question denominator.
Compared versions
3 profiles · recorded model, version, and settings
Compared versions
3 profiles · recorded model, version, and settings| Profile | Version or model | Configuration |
|---|---|---|
| OpenViking | 0.5.0 | Settings
|
| RAGFlow nightly | v1.0.0-rc2-48-g7eb673bfc | Settings
|
| RAGFlow rc1 | 1.0.0-rc1 | Settings
|