Tag
Google Research introduces a 'sleep framework' for LLMs that alternates active wake phases with offline sleep phases to consolidate short-term context knowledge into stable long-term parameters, combating catastrophic forgetting and enabling retention of up to 10M tokens of long-context knowledge.
论文提出 Cortex 方法，通过生成前瞻性查询并将上下文条件化输出蒸馏为可合并的低秩或结构化前馈网络更新，实现将上下文增益固化到参数记忆中，在保留 99.8% 原有非目标任务性能的同时恢复 88.9% 的上下文增益。
AutoViewMem is a data-driven framework that organizes long-term conversational memory into self-configuring semantic views to improve retrieval and personalization for LLM agents, showing superior performance on benchmarks.
The article details a 25-week experiment with an AI agent's memory system, using episodic and semantic memory with nightly consolidation, avoiding LLMs in memory promotion, and employing four separate graphs for association, evidence, causation, and temporal changes, while highlighting defects like attractor basins that degrade retrieval.
MemoryLACE is a lightweight memory framework for long-term LLM agents that explicitly models the lifecycle of textual evidence to improve reasoning without comprehensive knowledge graphs, achieving higher performance and reduced runtime on benchmarks.
This paper proposes 4MAS, a novel neural architecture inspired by biological bilaterality and memory consolidation, to address catastrophic forgetting in lifelong learning, achieving competitive results on benchmark datasets.
LycheeMemory V2 introduces semantic segment-level consolidation for LLM agent memory, reducing construction costs while improving or maintaining accuracy on LoCoMo and LongMemEval benchmarks.
This paper proposes that AI memory consolidation should recombine knowledge across domains (like dreaming) rather than merely replaying experiences, and demonstrates that cross-domain consolidation improves performance in both neural (LoRA fine-tuning) and symbolic systems.
This paper proposes using intrinsic device noise on analog neuromorphic hardware as a resource for continual learning by conditioning each weight's stochastic dynamics to avoid crossing memory-critical barriers, demonstrating non-monotonic retention improvement and validation on BrainScaleS-2 silicon.
Proposes a cognitively grounded multi-factor value function for agentic memory in LLM agents, learning interpretable weights to decide what to encode, forget, and retrieve under memory constraints. Improves gold-evidence retention significantly over similarity-only or recency-based baselines.
FadeMem introduces a distance-aware key-value memory consolidation mechanism that organizes historical video data into a temporal hierarchy, improving long-video generation under fixed cache constraints.
This paper introduces a 'Sleep' paradigm for large language models that enables continual learning through memory consolidation and dreaming phases, allowing models to distill short-term knowledge into long-term parameters and self-improve without human supervision.
This paper introduces a sleep-like consolidation mechanism for Transformer-based LLMs that periodically converts recent context into persistent fast weights in SSM blocks, clearing the KV cache to improve long-horizon reasoning without increasing inference latency.
Auto-Dreamer introduces a learned offline memory consolidation method for language agents, decoupling fast memory acquisition from slow cross-session consolidation, and achieving higher performance with smaller memory banks, generalizing to unseen environments.