A-Mem: Agentic Memory for LLM Agents

A Zettelkasten is a note-taking method: write small, self-contained notes and link them to related ones, so a network of ideas grows over time. Wujiang Xu, Yongfeng Zhang and colleagues at Rutgers and the AIOS Foundation built A-MEM to give LLM agents memory organized the same way, with the model doing the organizing.

Every interaction becomes a note. It holds the original content and a timestamp, plus attributes the model writes: keywords, tags and a context description that explains what the interaction means. An embedding of all of these is computed for search. When a note is added, A-MEM retrieves the most similar existing notes and asks the model which of them it should link to, based on shared attributes. Then comes the distinctive step, memory evolution: for each of those neighbors, the model decides whether to update its context, keywords and tags in light of the new note, and the updated version replaces the old one. At question time, the query is embedded and the top-k most similar notes go into the prompt.

The authors tested A-MEM on LoCoMo, long multi-session conversations of about 9,000 tokens with 7,512 questions across five types (single-hop, multi-hop, temporal, open-domain, adversarial), with six models from GPT-4o-mini to small Qwen and Llama models, and on DialSim, questions over TV-show dialogues. Compared with putting the whole conversation in context (the LoCoMo baseline), MemGPT, ReadAgent and MemoryBank, A-MEM had the best average rank and used far fewer context tokens per question. Its biggest gains were on temporal questions. On several question types, the full-conversation baseline or MemGPT scored higher.

An ablation showed that removing link generation and evolution hurt substantially, and removing only evolution hurt less.

A-MEM is one of several "smart memory" designs customers will ask about. It shows two useful ideas and one risky one.

Enrich at write time. Having the model write keywords, tags and a short context for each memory makes later retrieval much better than embedding raw text alone, especially for questions phrased differently from the original conversation.

Small retrieved context saves money. Answering from a handful of notes instead of the whole history cut context by about 85 to 93%. For long-running assistants, that is the difference between a viable and an unviable unit cost, provided writes are cheap enough.

In-place evolution needs history. Rewriting old notes when new information arrives can add useful connections, but it also changes what you stored, with no record of the original. For customer data, keep the original and store evolutions as versions you can audit and roll back.

Check per-type results, not averages. The paper's summary ranking hides that a simple full-context baseline beat A-MEM on some question types. Evaluate on the question mix your customer will actually ask.