Papers

Shared by Pavel · Updated 2026-11-12

The papers span frameworks, memory systems, benchmark design, and evidence about what memory should preserve.

Frameworks and memory systems

Paper and authorsMethod and contribution
CoALA: Cognitive Architectures for Language Agents (2023), Sumers et al.Frames agents as modular working, episodic, semantic and procedural memories, with structured actions and a repeated decision cycle including retrieval, reasoning and learning. It gives a vocabulary and architecture for comparing agent-memory designs.
MemGPT: Towards LLMs as Operating Systems (2023), Packer et al.Uses virtual context management: the agent moves information between fast in-context and slower external memory using tools and interrupts. It treats context limits as a memory-management problem.
HippoRAG (2024), Gutiérrez et al.Builds a knowledge graph with an LLM and uses Personalized PageRank for retrieval, drawing inspiration from hippocampal indexing. It offers graph-based retrieval for connected information.
Zep: A Temporal Knowledge Graph Architecture for Agent Memory (2025), Rasmussen, Paliychuk, Beauvais, Ryan and ChalefDescribes Graphiti, which combines conversation and business data in a temporal knowledge graph. The authors report 94.8% on Deep Memory Retrieval versus 93.4% for MemGPT, and up to 18.5% accuracy gain with 90% lower latency on LongMemEval versus baselines. These are the paper’s reported results.
Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory (2025), Chhikara et al.Extracts candidate facts from new messages, retrieves similar memories, then chooses ADD, UPDATE, DELETE or NOOP. The paper evaluates its memory variants on benchmarks including LoCoMo and LongMemEval.
A-MEM: Agentic Memory for LLM Agents (2025), Xu et al.Turns each new memory into a structured note with contextual descriptions, keywords and tags, links related notes, and allows new notes to update older ones. It adapts linked-note organization to evolving agent memory.
Rethinking How to Remember: Beyond Atomic Facts in Lifelong LLM Agent Memory (TriMem) (2026), Sun, Zhu, Yao, Liu, HanKeeps raw dialogue segments linked to source IDs alongside atomic facts and synthesized profiles. It is a hybrid design that retains source material while providing structured abstractions.

Benchmarks and evaluation

Paper and authorsMethod and contribution
LoCoMo (2024), Maharana et al.Tests question answering, event summarization and multimodal dialogue over long conversations—averaging 300 turns and 9K tokens, with up to 35 sessions. Its reported answer-key errors are a caveat when interpreting scores.
LongMemEval (2024; ICLR 2025), Wu et al.Uses 500 questions to test information extraction, multi-session and temporal reasoning, knowledge updates, and abstention. The benchmark helps separate memory abilities, but scores across papers are usually not directly comparable.
MemoryAgentBench (2025; ICLR 2026), Hu, Wang and McAuleyEvaluates accurate retrieval, test-time learning, long-range understanding and selective forgetting in an incremental multi-turn setup, reformulating existing datasets and adding EventQA and FactConsolidation.

Source fidelity and episodic memory

Paper and authorsMethod and contribution
Fidelity Before Structure: Verbatim Chunks Beat Lossy Artifact Extraction in Long-Conversation LLM Memory (2026), Tao AnIn a single-author ablation holding the model, retriever, reranker and judge fixed, verbatim chunks scored 43.9% vs 28.0% on LoCoMo and 67.4% vs 45.4% on LongMemEval-S. A one-hop graph did not close the gap; chunks abstained worse. The paper recommends adding artifacts beside source chunks, not replacing them.
Episodic Memory Is the Missing Piece for Long-Term LLM Agents (2025), Pink et al.A position paper that argues episodic memory—remembering specific experiences with their time and context—belongs in the research vocabulary for agent memory. It is conceptual, not a benchmark result.