Tag
论文提出ConMem算法，将LLM智能体的长期记忆保留决策重构为反事实因果推断，并构建MemAlign基准（1,079个记忆案例）来研究记忆错位问题，实验表明该方法在缓解过度保留偏差的同时保持个性化效用。
The MemTensor team releases MemT 1.0, pushing memory mechanisms from bolt-on RAG and Agent Memory into the core of model computation, allowing models to accumulate experience while solving tasks. Official tests show improvements of over 5% in Coding and Working scenarios, and the source code is now open-sourced on Hugging Face.
A HuggingFace article summarizing a paper by IIT Madras researchers that groups LLM memory handling into five approaches — store more, compress, correct, learn at inference, and distribute plasticity — and questions how they compare when new information conflicts with old state.
The paper introduces Memorilla, a memory system that compresses an entire document collection into 16 query-dependent memory tokens for frozen LLMs, outperforming full-context inference on FactKG (84.33% vs 80.43%) while using 108x fewer tokens and beating baselines like Mem0 and MemoRAG across five of eight benchmarks.
提出 T2Mem 框架，将记忆构建延迟到推理阶段，随推理过程渐进式构建面向查询的记忆，在 LoCoMo 和 LongMemEval 上超越强基线并显著降低记忆构建开销。
Jev-Mem introduces an agentic memory architecture inspired by System-One/System-Two cognition, enhancing efficiency and effectiveness for long-horizon AI agents with improved scores and faster operations.
This paper introduces Fortunate Recall, an ontology-driven policy layer for managing memory lifecycle in LLMs, which classifies personal facts into behavioral categories and applies differential policies to improve coherence and reduce confabulation, outperforming existing systems on benchmarks.
The author developed a Datalog-based memory system for LLMs to maintain accurate state during investigations like vulnerability research, automatically updating conclusions when facts change.
The article introduces RENDER, a benchmark that controls reader-facing evidence in LLM memory evaluations, showing significant accuracy variations based on how memory entries are rendered to the model.
The author argues that for users who already subscribe to LLM services like ChatGPT Pro, running local embedding and reranker models for a memory system is more practical than running local LLMs, and details their GBrain-based setup.
Mnemo is an open-source, local-first memory layer for any LLM that extracts entities and relationships into a persistent knowledge graph using SQLite and petgraph, providing automatic context injection for enhanced conversations.
SAGE proposes a novelty gate for memory evolution in agentic LLMs, using a von Mises-Fisher-based density estimator to decide whether to add, merge, or ignore new facts, reducing LLM calls while maintaining memory quality.
Introduces SeqMem-Eval, a diagnostic evaluation framework for sequentially evolving LLM memory that measures multiple dimensions beyond aggregate metrics, revealing trade-offs between adaptability and stability.
GroupMemBench is a new benchmark for evaluating LLM agent memory in multi-party conversations, exposing failures in current memory systems with the best achieving only 46% average accuracy.
This research demonstrates that continuously updating LLM agent memories through distillation and consolidation loops causes performance regression, even when trained on ground-truth solutions. The study finds that episodic-only retention outperforms text-based consolidation, highlighting significant flaws in current self-improvement paradigms.
An independent developer quietly shipped a working “LLM Wiki” system—persistent memory for language models—weeks before Andrej Karpathy publicized the same concept.
Someone implemented a working "LLM Wiki" system a month before Andrej Karpathy publicized the concept, addressing the problem that LLMs restart from zero without memory or learning.
Researchers introduce BEHEMOTH benchmark and CluE cluster-based prompt optimization to enable LLMs to extract and retain heterogeneous memory across diverse tasks, achieving 9% gains over prior self-evolving frameworks.
Mem0 introduces a scalable memory-centric architecture using graph-based representations to improve long-term conversational coherence in LLMs, significantly reducing latency and token costs while outperforming existing memory systems.