Dhravya Shah announced that Supermemory achieved ~99% accuracy on the LongMemEval benchmark using a new experimental technique called ASMR (Agentic Search and Memory Retrieval). Importantly, Shah explicitly framed this as a "stunt" or social experiment—not their production system—to make a point about how memory systems should be evaluated. The experimental architecture uses multi-agent orchestration and active agentic reasoning instead of traditional vector embeddings and RAG approaches. This represents a significant improvement over their previous ~85.86% benchmark result and outperforms all publicly benchmarked memory systems. The team plans to open-source the experimental code in early April 2026.
The announcement sparked significant community attention, with Shah later clarifying that the experiment was designed as a parody to establish new standards for reporting memory system quality. The work demonstrates key engineering insights: agentic retrieval outperforms vector search for temporal data, parallel processing is critical for performance, and specialized agents beat generalized prompts.
Key Takeaways
ASMR (Agentic Search and Memory Retrieval) achieves ~99% on LongMemEval by replacing vector embeddings with multi-agent orchestrated reasoning—a fundamental architectural shift away from traditional RAG
The experimental system uses specialized parallel agents (3 reader agents, 3 search agents) rather than a single generalized model, proving specialization significantly outperforms generalization
Explicitly framed as a 'stunt' and social experiment to create a new standard for how memory system quality should be reported in the AI industry
Demonstrates the theoretical ceiling of agent memory performance: improved from production system's ~85.86% to ~99% accuracy
Code and experimental architecture will be open-sourced in early April 2026 for community learning and adoption
Agentic retrieval eliminates the 'semantic similarity trap' of vector search when handling temporal data, contradictions, and knowledge updates in multi-session conversations
About
Author: Dhravya Shah (Founder, Supermemory)
Publication: X (Twitter) / Supermemory Blog
Published: 2026-03-21
Sentiment / Tone
Excited and tongue-in-cheek; confident but self-aware about the experimental nature. The post balances genuine technical achievement with playful acknowledgment that this was deliberately designed as a "stunt" to make a broader point about evaluation standards.
Dhravya Shah's Follow-up Clarification Post Follow-up X post clarifying that the announcement was a social experiment and parody to establish new standards for memory system evaluation
Research Notes
This announcement went viral in the AI community, with significant buzz around the ~99% benchmark score. However, Dhravya Shah's follow-up posts clarify that this was an intentional social experiment framed as a parody. The core innovation—using multi-agent orchestration instead of vector embeddings for memory retrieval—is technically sound and represents a genuine architectural advance. The achievement is meaningful not because it's their production system, but because it demonstrates the theoretical upper bound of what's possible with agent-based memory systems and provides proof-of-concept for techniques that could eventually be integrated into production systems. The decision to open-source the code in April 2026 demonstrates commitment to advancing the broader AI community's understanding of agent memory systems.
Topics
AI Memory SystemsAgent ArchitectureMulti-Agent SystemsLLM BenchmarkingRetrieval-Augmented Generation (RAG)Temporal ReasoningLongMemEval BenchmarkAI ResearchProduction-Grade AI Systems