AI News AI资讯 22h ago Updated 19h ago 更新于 19小时前 49

Red Bear AI Compleces Hundreds of Millions of Yuan A+ Round, Extending from To B Services to To C Applications Based on AI 'Memory Science' 红熊AI完成数亿元A+轮融资,基于AI“记忆科学”从To B服务延伸至To C应用|36氪首发

Hongxing AI secured hundreds of millions of RMB in A+ financing, reaching a post-money valuation of nearly 3 billion RMB, marking its sixth funding round in 15 months. The company’s core technology, "AI Memory Science," integrates the OpenBear MoE-based LLM with the MemoryBear system to solve long-context forgetting and high token costs. MemoryBear utilizes biological memory models (ACT-R, Ebbinghaus curve) to achieve <1% knowledge forgetting, 25x reduction in token consumption, and ~0.2% halluc 红熊AI完成数亿元A+轮融资,投后估值近30亿元,资金将用于深化“AI记忆科学”研究及拓展To C应用。 提出“AI记忆科学”概念,通过OpenBear大模型与MemoryBear记忆系统融合,解决大模型上下文遗忘、幻觉及Token成本高等痛点。 构建“模型+记忆+Agent”三角矩阵,实现知识遗忘率<1%、Token消耗降低25倍、幻觉率降至0.2%的技术指标。 商业化进展迅速,2026上半年确收超1.7亿元,ARR突破0.5亿元,客户超500家,并计划推出面向C端的通用大模型。

75
Hot 热度
70
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Hongxing AI secured hundreds of millions of RMB in A+ financing, reaching a post-money valuation of nearly 3 billion RMB, marking its sixth funding round in 15 months.
  • The company’s core technology, "AI Memory Science," integrates the OpenBear MoE-based LLM with the MemoryBear system to solve long-context forgetting and high token costs.
  • MemoryBear utilizes biological memory models (ACT-R, Ebbinghaus curve) to achieve <1% knowledge forgetting, 25x reduction in token consumption, and ~0.2% hallucination rates.
  • Commercially, Hongxing AI reported over 170 million RMB in confirmed revenue in H1 2026, with ARR exceeding 50 million RMB and over 500 enterprise clients.
  • The company is expanding from B2B services (customer service, marketing) to B2C applications, launching the OpenBear consumer model and CodeBear programming tool in July 2026.

Why It Matters

This case highlights the industry's pivot from competing solely on model parameters to focusing on application-layer stability, specifically through advanced memory management systems. For practitioners, it demonstrates that integrating dynamic, biologically-inspired memory architectures can significantly reduce operational costs (tokens) and improve reliability (hallucination reduction) in enterprise settings. The rapid commercialization and funding success signal that "memory-centric" AI agents are becoming a critical infrastructure layer for scalable AI deployment.

Technical Details

  • Architecture: Combines OpenBear, a trillion-parameter MoE (Mixture of Experts) model designed with native memory interfaces, with the proprietary MemoryBear system.
  • Memory Mechanisms: MemoryBear employs a layered dynamic structure (working, short-term, long-term) based on ACT-R dual-memory architecture and Ebbinghaus forgetting curves. Key technologies include time-memory tracking, dynamic semantic networking, and intelligent semantic pruning to remove redundant data.
  • Performance Metrics: Claims a knowledge forgetting rate of under 1%, a 25-fold reduction in token consumption, industry ambiguity preprocessing under 1%, and hallucination rates reduced to approximately 0.2% via a 3D self-reflection engine.
  • Agent Framework: Features a "native agent" architecture where agents possess built-in memory capabilities for continuous interaction history and task execution, supporting autonomous planning and multi-agent collaboration with minimal memory sharing.
  • Developer Tools: CodeBear is a memory-driven coding platform that accumulates project-specific knowledge and code norms across interactions to maintain context continuity in complex development tasks.

Industry Insight

  • Cost Efficiency via Memory Pruning: The significant reduction in token usage (25x) suggests that memory optimization is a viable strategy for lowering the total cost of ownership (TCO) for large-scale AI deployments, moving beyond simple model scaling.
  • Standardization of Memory Layers: As multiple startups (Engram, Clipto, MemOS) focus on AI memory, the industry is likely to see the emergence of standardized "memory operating systems" or APIs, making memory management a distinct layer in the AI stack separate from base models.
  • B2B to B2C Expansion Strategy: The move toward consumer-facing products indicates that robust memory capabilities are a key differentiator for personalized user experiences, suggesting future consumer AI tools will heavily rely on persistent, personalized memory profiles rather than stateless interactions.

TL;DR

  • 红熊AI完成数亿元A+轮融资,投后估值近30亿元,资金将用于深化“AI记忆科学”研究及拓展To C应用。
  • 提出“AI记忆科学”概念,通过OpenBear大模型与MemoryBear记忆系统融合,解决大模型上下文遗忘、幻觉及Token成本高等痛点。
  • 构建“模型+记忆+Agent”三角矩阵,实现知识遗忘率<1%、Token消耗降低25倍、幻觉率降至0.2%的技术指标。
  • 商业化进展迅速,2026上半年确收超1.7亿元,ARR突破0.5亿元,客户超500家,并计划推出面向C端的通用大模型。

为什么值得看

本文揭示了AI行业从“拼参数”向“拼应用与记忆层”转型的关键趋势,展示了如何通过类脑记忆机制解决大模型落地的核心痛点。对于从业者而言,红熊AI在智能客服、营销等领域的量化成效及快速商业化路径,为AI Agent的记忆增强技术提供了极具参考价值的实战案例。

技术解析

  • OpenBear通用大模型:采用MoE稀疏混合专家架构,拥有百万亿级参数。设计上专为记忆优化,内置记忆接口与注意力机制优化模块,支持高频低延迟的信息交互,强调跨领域泛化与自我反思能力。
  • MemoryBear记忆科学系统:借鉴人类认知机制,融合ACT-R双记忆架构、艾宾浩斯遗忘曲线等理论。包含工作、短期、长期三层动态记忆体系,核心技术包括时间记忆、动态语义网络、智能语义剪枝(降低冗余)、3D自我反思引擎(修正幻觉)及最小化记忆共享技术。
  • 原生Agent架构:具备“记忆原生”特性,内置记忆能力以记录交互细节与任务结果,支持自主规划与多Agent协同,实现动态任务执行与自动续传,推动AI从被动工具向主动协作演进。
  • 性能指标:在MemoryBear支持下,基础大模型知识遗忘率控制在1%以下,Token消耗降低25倍,行业歧义预处理率<1%,模型幻觉率降至0.2%左右。

行业启示

  • 记忆层成为大模型落地新基建:随着算力竞赛放缓,解决上下文窗口限制、记忆一致性及成本控制问题成为关键,“记忆科学”或将成为下一代AI应用的核心竞争力。
  • 商业化验证加速:红熊AI在一年内实现从融资到大规模营收的跨越,表明具备明确痛点解决方案(如客服、营销)的垂直领域AI应用已具备成熟的商业闭环能力。
  • To B向To C延伸趋势:头部AI应用厂商开始尝试将经过B端验证的记忆技术下放至C端通用大模型及开发者工具(如CodeBear),预示着AI个人助理与开发提效工具的爆发期临近。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Funding 融资 Product Launch 产品发布 Conversational AI 对话系统 Education AI 教育AI