Scholar Feed

Agent / long-term memory

Memory-R1

Memory-R1: Enhancing Large Language Model Agents to Manage and Utilize Memories via Reinforcement Learning

Superseded baseline#12 of 63 most-superseded · first seen Aug 27, 2025

Superseded — cited as a baseline and beaten by newer methods

4 papers critique it · 1 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites Memory-R1 as a baseline.

Memory-R1 introduces a comprehensive forgetting mechanism similar to the one we propose, but lacks any kind of temporal encoding in its stored memories.
— Mnemosyne: An Unsupervised, Human-Inspired Long-Term Memory Architecture for Edge-Based LLMs
“While Memory-R1~memoryr1 advances this by using reinforcement learning to optimize storage and retrieval policies, its retrieval is still a single-round process.”
— General Agentic Memory Via Deep Research
“In long-term dialogues, a single reward given at the end of a multi-session trajectory is too sparse.”
— MemBuilder: Reinforcing LLMs for Long-Term Memory Construction via Attributed Dense Rewards
“However, they rely mainly on outcome-level rewards and do not explicitly address cross-session credit assignment under diverging memory states.”
— Memory-R2: Fair Credit Assignment for Long-Horizon Memory-Augmented LLM Agents

Beaten on benchmarks

Head-to-head results where a newer method reports beating Memory-R1. Values are copied from the source paper's tables — verify against the cited paper.

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.