FIND-20260323-028 · 2026-03-23 · Innovation Veille
Supermemory ASMR: ~99% SOTA on LongMemEval — Persistent Memory Layer for AI Agents
adhoc
HIGH
Dhravya Shah (founder of Supermemory) announced a breakthrough in agent memory: the ASMR (Agentic Search and Memory Retrieval) architecture achieves ~99% accuracy on LongMemEval_s, the standard benchmark for long-term conversational memory. The approach uses parallel reader agents, active agentic retrieval (reasoning over stored facts rather than pure vector search), and multi-variant ensembles running 8 specialized prompt variants in parallel. Supermemory already ships a production Memory API, a Claude Code plugin (supermemoryai/claude-supermemory, 2.4k stars, MIT), an OpenCode plugin, and MCP v4 integration. Open-source code for the ASMR system is committed for early April 2026.
Source
https://x.com/dhravyashah/status/2035517012647272689
ODS Impact
Directly relevant to ADLC pipeline agent memory: the Claude Code plugin is a drop-in for persistent context across agent sessions, already used in this environment. The Memory API (hosted, MIT) could power cross-agent shared context for ODS pipeline agents — BA, architect, dev, security agents currently lose state between runs. The Conversations endpoint (processes only new tokens, reduces cost) and Hybrid Search feature are relevant for any ODS service requiring RAG or chatbot functionality. The code-chunk npm package (28-point recall improvement via AST-aware splitting) is relevant to pdf-engine and docstore search indexing. The ASMR open-source release in April would allow self-hosting a high-accuracy memory backend on ODS infrastructure.
Security Review
License: MIT | Maintenance: ACTIVE | Risk: MEDIUM | Recommendation: USE_WITH_CAUTION
Tags
ai-memory
agents
mcp
claude-code
rag
typescript
open-source
benchmark