memory retrieval
Beyond Market Intelligence keeps memory retrieval in one place: 2 stories so far. The section currently leads with “96% accuracy on LongMemEval-S: Our memory retrieval outperforms multi-turn agents” and “A New Protocol for Measuring How Conversational Memory Evolves Over Time”. 96% accuracy on LongMemEval-S is a strong result, but the real story is how it was achieved. Measuring how conversational memory evolves over time requires more than endpoint questions. Acme AI is the next-generation, AI-powered spreadsheet platform built to replace Excel and redefine how analysts, data scientists, and enterprise teams work… The list below is every memory retrieval story on Beyond Market Intelligence, newest first.
96% accuracy on LongMemEval-S: Our memory retrieval outperforms multi-turn agents
96% accuracy on LongMemEval-S is a strong result, but the real story is how it was achieved. Papr's memory retrieval hit that score using a 0.6B open-source embedding model and a single-turn answer model, not complex multi-turn agents. The team's transparency is refreshing, they acknowledge the top-20 tuning and note this run was optimized for token efficiency. If you want to see how simpler approaches can outperform more elaborate systems, this is worth exploring.

A New Protocol for Measuring How Conversational Memory Evolves Over Time
Measuring how conversational memory evolves over time requires more than endpoint questions. LSREP, a new protocol from independent researcher Deepnar S, evaluates whether memory systems genuinely maintain evolving state across 1,985 turns. The findings are honest: ICE v2 didn't outperform vector-RAG. That's the value here. The methodology matters more than the result. If you appreciate rigorous evaluation design, "How to Build Fair Evaluation Sets with Combinatorial Precision" offers complementary thinking on constructing balanced tests.