9 papers

Memory Systems

How agents retain, recall, and reflect across long-running work.

Papers
9
Read time
2.3h
Avg length
15m
01
Memory Systems

HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models

Use an LLM to turn your whole corpus into one big knowledge graph, then answer multi-hop questions in a *single* retrieval step by running Personalized PageRank from the query's entities — instead of paying for 3-5 round trips of iterative LLM-driven retrieval.

BERNAL JIMÉNEZ GUTIÉRREZ, YU GU, YIHENG SHU, MICHIHIRO YASUNAGA, YU SU (OSU + STANFORD) · NEURIPS 2024
Read
16 min
Source
arXiv:2405.14831
02
Memory Systems

MemGPT: Towards LLMs as Operating Systems

Treat the LLM's context window like RAM and give the model OS-style function calls to page its own memory in and out of an external database, so a fixed-context model behaves as if it had unbounded memory.

CHARLES PACKER, SARAH WOODERS, KEVIN LIN, VIVIAN FANG, SHISHIR G. PATIL, ION STOICA, JOSEPH E. GONZALEZ (UC BERKELEY) · 2024
Read
14 min
Source
arXiv:2310.08560
03
Memory Systems

From RAG to Memory: Non-Parametric Continual Learning for Large Language Models

HippoRAG 2 turns a pile of documents into a queryable "long-term memory" by building a knowledge graph over passages *and* concepts, then using Personalized PageRank to retrieve — beating plain vector search on multi-hop questions without breaking on simple ones.

BERNAL JIMÉNEZ GUTIÉRREZ, YIHENG SHU, WEIJIAN QI, SIZHE ZHOU, YU SU (OHIO STATE / UIUC) · ICML 2025
Read
14 min
Source
arXiv:2502.14802
04
Memory Systems

Memento: Fine-tuning LLM Agents without Fine-tuning LLMs

Instead of retraining the LLM, Memento gives an agent a growing "case bank" of its own past successes and failures and learns a tiny side-model that decides which past cases to recall — so the agent keeps getting better on the job without ever touching the model weights.

ZHOU, CHEN, GUO, YAN ET AL. (UCL, HUAWEI NOAH'S ARK, SUSTECH) · 2025
Read
16 min
Source
arXiv:2508.16153
05
Memory Systems

ArcMemo: Abstract Reasoning Composition with Lifelong LLM Memory

Instead of caching whole question/answer pairs, ArcMemo distills each solved problem into small, reusable, *named* concepts in plain language, then selectively pulls the relevant ones into the prompt for the next problem — giving an LLM a growing, composable skill library without any fine-tuning.

MATTHEW HO, CHEN SI, ZHAOXIANG FENG, FANGXU YU, YICHI YANG, ZHIJIAN LIU, ZHITING HU, LIANHUI QIN (UC SAN DIEGO, U. MARYLAND)
Read
16 min
Source
arXiv:2509.04439
06
Memory Systems

FERNme — Action-Coupled, Cost-Bounded Memory for Multi-Tenant Agents

Stop paying an LLM to "remember" each user on every turn — instead keep each user

MIRKOMIL SHARIPOV (ACQUILAB INC.) · PREPRINT V0.1.0, 2026 · GITHUB.COM/MIRKOFR/FERNME
Read
11 min
07
Memory Systems

Hindsight is 20/20: Building Agent Memory that Retains, Recalls, and Reflects

A memory architecture that stops treating agent memory as a dumb top-k vector store and instead splits it into four "networks" (facts vs. experiences vs. opinions vs. summaries) with three operations (retain / recall / reflect), lifting a 20B open model from 39% to 83.6% on long-horizon memory benchmarks — beating full-context GPT-4o.

LATIMER, BOSCHI, NEESER, BARTHOLOMEW, SRIVASTAVA, WANG & RAMAKRISHNAN (VECTORIZE.IO / WASHINGTON POST / VIRGINIA TECH)
Read
18 min
Source
arXiv:2512.12818
08
Memory Systems

MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents

Instead of letting an agent's context balloon as it works through a long, multi-step task, MEM1 trains the model with RL to rewrite a single compact "internal state" each turn — folding new observations into it and throwing everything else away — so memory stays roughly constant no matter how long the task runs.

ZHOU, QU, WU, KIM, PRAKASH, RUS, ZHAO, LOW, LIANG (SMART · NUS · MIT · YONSEI)
Read
16 min
Source
arXiv:2506.15841
09
Memory Systems

MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent

Instead of stretching a model's context window, MemAgent teaches an LLM (via RL) to read a giant document in small chunks while continuously rewriting a fixed-size scratchpad of notes — so an 8K-window model trained on 32K text answers questions over 3.5M-token documents with under 5% accuracy loss and linear cost.

HONGLI YU, TINGHONG CHEN, JIANGTAO FENG, ET AL. (BYTEDANCE SEED · TSINGHUA AIR)
Read
18 min
Source
arXiv:2507.02259