Thalamus AI Secures Seed Funding for Multimodal Long-Term Memory in AI

Thalamus AI, a startup focused on multimodal long-term memory for artificial intelligence, has successfully completed a seed funding round of tens of millions of yuan. The investment, from Shenzhen-based funds and industry capital, will support technological research and development, as well as talent acquisition. Founded in November 2025, Thalamus AI aims to bridge the gap towards 'proactive intelligence' by enabling AI systems to understand users deeply and interact proactively. The company's core offering is its native multimodal memory base, MemAura, which addresses key challenges in the current AI memory landscape. These challenges include high inference costs associated with full context window utilization, context fragmentation across sessions and tasks, and the limitation of existing solutions to purely text-based data, neglecting multimodal inputs like vision and audio. MemAura is designed to support long-term memory accumulation and rapid retrieval, employing an efficient context mechanism to reduce costs. It boasts a 40%-49% reduction in token consumption, memory retrieval latency under 400 milliseconds, and initial response times within one second, with an overall accuracy exceeding 80%. The system mimics human memory by identifying key information, filtering noise, encoding data, and partitioning private and non-private memories. It also utilizes a hot and cold storage strategy for efficient data access and cost control. Thalamus AI's founder, Zhang Yuan, believes that a dedicated, independent memory layer is crucial for AI to evolve from general-purpose to personalized and ultimately proactive intelligence, citing the issue of 'memory islands' created by users switching between different foundation models. The company plans to offer its services through an ADK (Agent Development Kit) and API, with current clients primarily in companion hardware, AI customer service, and digital employee sectors. Zhang anticipates a surge in demand for multimodal long-term memory as AI integrates more deeply with the physical world and embodied AI applications mature.