LeanMem splits AI memory by type, gains 15.1 points

LeanMem sorts every piece of a conversation into one of three memory types, then stores each differently, per a paper posted to arXiv.

Most memory systems for AI agents run everything through one summarization process. That wastes tokens compressing facts that will never change, and loses detail on facts that need to stay exact. LeanMem beat the strongest baseline in every setting tested on LoCoMo and LongMemEval-S, gaining up to 15.1 points. It also ran at the lowest or near-lowest inference token count and latency of the methods compared.

The three types split by how a piece of dialogue behaves over time. Stable facts about a person go into compact profile memory. Things that change as the conversation evolves become temporally structured event memory. Anything that needs to stay traceable to its exact source, like quotes, numbers, or commitments, gets stored as source-grounded record memory.

What to copy

The part worth taking for production agents is the maintenance rule, not the classification scheme. LeanMem only re-processes event memory when it updates. Profile and record memory get written once and left alone. Most memory implementations re-summarize the whole store on every update cycle, which is where the token cost actually comes from.

The results held across two different models, GPT-4.1-mini and Qwen3-8B, which suggests the gain comes from the architecture and not from either model's quirks.

The paper doesn't say how the routing decision gets made at write time, whether by a classifier, a prompt, or hand-coded rules. That's the detail that decides whether the approach is easy to reproduce.