Tag
作者声称其自研的混合架构在 ARC-AGI-3 基准上取得 15.85%（29/183 关卡）的成绩，约 4 倍于 2026 里程碑获奖水平。该架构由稀疏化的 Gemma-4-31B-It 结合 OpenJEV、HRM/TRM「右半球」及双向胼胝体子系统构成，并在 QASPer LongBench 上以 62.23% KV cache 达到接近稠密模型的性能，附两篇预印本和开源仓库。
该论文提出了一种由离散化偏微分方程驱动的空间场替代传统神经元间通信的二阶循环模型，实现等价于无限阶RNN的无界感受野，并通过边缘稳定性约束消除梯度消失与爆炸问题，在长序列基准上以更少参数超越其他循环模型。
CharacterMemory is an open-source memory system for AI characters inspired by neuroscience, retrieving information based on emotion, recency and connections, with world/emotion tracking, routine and calendar management, a WebUI, and easy integration into existing code.
This paper proposes a Dual-Layer Agentic Memory framework for LLM agents that manages knowledge lifecycle through fast write routing and slow consolidation to prune redundant external memory while maintaining performance.
Proposes a reinforcement learning framework that uses locally linear embeddings to capture environment dynamics and an attention mechanism to adaptively fuse dynamics-specific and reward-specific features, inspired by neural principles, improving learning efficiency.
The article describes Skynet, an Elixir-based framework using OTP GenServers to build persistent cognitive architectures for LLM agents. It implements a layered memory stack inspired by neuroscience, addressing the amnesia problem in long-running agents.
Microsoft researchers propose a biologically-inspired memory architecture for LLM agents that incorporates mechanisms like sleep-phase consolidation and interference-based forgetting to manage persistent memory efficiently.
The paper introduces Mela, a memory-augmented transformer architecture inspired by human memory consolidation, featuring a Hierarchical Memory Module that improves long-context language modeling performance.