Thalamic Intelligence: China's Only Multimodal Long-Memory Company Raises Tens of Millions in Seed Funding, Betting on Proactive AI

Thalamic Intelligence, founded in November 2025 by CEO Zhang Yuan, a Peking University graduate with dual degrees in electronics and economics, has raised a seed round of tens of millions of yuan from a Shenzhen first-tier fund and industrial capital, to be used for R&D and talent expansion. The team, averaging about 26 years old, hails from institutions like Alibaba DAMO Academy, Tencent, SenseTime, CUHK, and Peking University. The company positions memory as an independent infrastructure layer, arguing that cross-model memory silos necessitate a third-party, user-centric memory layer. Its product, MemAura, launched in June, is a native multimodal memory base that offers ADK and API services, featuring a bionic memory processing mechanism that mimics human memory: event recognition, encoding, and privacy partitioning, along with hot/cold storage strategies. Performance metrics include a 40%-49% reduction in input token consumption, memory retrieval latency under 400 milliseconds, end-to-end first response under 1 second, and overall accuracy above 80%. The company's first-generation STKG spatiotemporal knowledge graph achieved SOTA on LoCoMo and LongMemEval benchmarks. In May 2026, Thalamic Intelligence, with NVIDIA, HKUST, and CUHK, released MEMLENS, the world's first multimodal long-memory benchmark, open-sourced and ranked in the top three of Hugging Face Daily Papers. The benchmark revealed that both visual language models and memory agents suffer performance degradation and increased hallucination with longer contexts, and that some post-training methods weaken refusal abilities. The company targets companion hardware and vertical Agent scenarios, charging via API calls and scenario licenses. CEO Zhang Yuan believes the endgame for memory is to support proactive AI, where AI initiates interactions based on deep user understanding, and that memory's core lies in decision-making, not storage. She also predicts a surge in embodied memory needs for home robots. The company is exploring E2P techniques and transformer fine-tuning for next iterations.