# ASMR Technique

> Agentic Search and Memory Retrieval technique by Supermemory using parallel reader agents and search agents that achieved ~99% accuracy on LongMemEval benchmark.

- **URL:** https://blog.supermemory.ai/we-broke-the-frontier-in-agent-memory-introducing-99-sota-memory-system/
- **Category:** Concepts & Definitions
- **Tags:** agent-memory, retrieval, multi-agent
- **Featured:** yes
- **Updated:** 2026-03-24 13:36
- **Canonical page:** https://vectordb.ever.works/items/asmr-technique

## Details

## Overview

ASMR (Agentic Search and Memory Retrieval) is a breakthrough technique developed by Supermemory that uses multi-agent orchestration instead of traditional vector database retrieval.

## Architecture

### Parallel Reader Agents
The system deploys an agent orchestrator utilizing 3 parallel reader (observer) agents powered by Gemini 2.0 Flash, which read through raw sessions concurrently. Their goal is targeted knowledge extraction across six vectors:
- Personal Information
- Preferences
- Events
- Temporal Data
- Updates
- Assistant Info

### Parallel Search Agents
When a question arrives, the system does not query a vector database but instead deploys 3 parallel search agents:
- Agent 1: Searches for direct facts and explicit statements
- Agent 2: Looks for related context, social cues, and implications
- Agent 3: Provides additional perspectives

## Performance

ASMR achieved approximately 99% on LongMemEval_s, published on March 22, 2026, putting it ahead of every publicly benchmarked memory system.

## Why It Works

By using agents that actively read and reason over stored findings rather than relying on semantic similarity alone, ASMR better handles:
- Contradictory information
- Temporal reasoning
- Multi-session context
- Complex queries requiring synthesis

## Pricing

Planned for open-source release in April 2026.

---

_This Markdown mirror is generated by Ever Works for AI agents. The canonical HTML page is at https://vectordb.ever.works/items/asmr-technique._