We broke the frontier in agent memory: To prove a point.

Supermemory presents ASMR, an experimental multi-agent memory-retrieval architecture that it claims achieved roughly 99% accuracy on the LongMemEval-s benchmark, although the post later states that the announcement was a parody and social experiment intended to encourage better standards for evaluating memory systems. LongMemEval tests long-term AI memory across large, multi-session conversation histories containing conflicting, updated, and temporally distributed information, where retrieval noise and outdated facts often limit performance. The proposed system replaces conventional vector-database retrieval with parallel reader agents that extract structured facts from sessions, search agents that identify direct evidence, contextual implications, and timelines, and specialized answer-generating agents that evaluate retrieved context. Two reported approaches included an eight-prompt ensemble scoring 98.6% when any variant found the correct answer and a 12-agent decision forest with an aggregator model producing a single consensus answer at 97.2%. The authors argue that agentic retrieval, parallel processing, and specialized reasoning can outperform general-purpose RAG approaches for temporal memory tasks, and they say they plan to open-source the experimental implementation while exploring how such methods could be adapted to production systems.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.