Breaking

A-MEM Released: Self-Evolving Memory for LLM Agents Built on Zettelkasten

July 2, 2026 at 14:11 EDT

AI Agents
Open Source
Research & Papers

A-MEM (Agentic Memory for LLM Agents), a dynamic, self-evolving memory system for LLM agents, has been released as open source on GitHub and is drawing developer attention. Built on the principles of the German knowledge-management method "Zettelkasten," its distinguishing feature is that the agent itself generates, links, and evolves its memories.

NeurIPS 2025 Poster · Rutgers-led research team

A-MEM: an agent memory that writes, links, and evolves its own notes

Inspired by the Zettelkasten method, A-MEM lets an LLM autonomously build atomic memory notes, connect them by meaning, and update older memories as new ones arrive — outperforming fixed-structure memory systems on long-term dialogue.

44.27

Multi-hop ROUGE-L (GPT-4o-mini) — more than double the 18.09 baseline

+192%

DialSim F1 vs. MemGPT (+35% vs. LoCoMo baseline)

Foundation models tested — beats SOTA baselines across all

Multi-hop reasoning: A-MEM vs. LoCoMo baseline

ROUGE-L on GPT-4o-mini — higher is better

18.09

LoCoMo
baseline

44.27

A-MEM

How a memory is formed and kept alive

Note Construction

LLM writes an atomic note with context, keywords & tags

→

Link Generation

Connects to past notes by semantic similarity

→

Memory Evolution

New notes update the context & attributes of older ones

Strengths

Dynamic Zettelkasten-style graph for long-term tasks
More adaptable than fixed-structure memory
Strong on complex, multi-hop reasoning
Resource-efficient design

Open challenges

LLM-driven memory adds API cost & latency
Memory evolution can be unpredictable
Benchmarks are conversation-centric
Complex implementation, limited real-world validation

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.