Long-context retrieval
Recall Core should improve how relevant evidence is found, ranked, compressed, and carried into the next model call.
Research
Memory quality should be measurable, not just marketed.
Recall Core research tracks retrieval, context selection, grounding, and long-running memory behavior while keeping improvements general enough for real users.
Re:Yi official sitePublic content, local custody.Recall Core data stays outside the online service by default.
Signal
Recall Core retrieval behavior
Signal
Context assembly and compression
Signal
Contradiction and time sensitivity
Signal
No benchmark-specific prompt patches
Grounded answer packaging
The system should preserve source traceability and avoid hiding uncertain or missing memory behind confident prose.
General capability first
Benchmark work must not become question-specific shortcuts. Fixes should improve ordinary users and enterprise cases too.
Benchmark discipline
Every research improvement must survive normal product use.
LongMemEval-style work is useful only when it improves retrieval, ranking, context packaging, and uncertainty handling for general memory users. Public pages should avoid unsupported leaderboard claims.