Xisiid Intelligence has raised RMB 25 million in seed funding to develop a brain-inspired AI architecture designed to help agents execute complex, long-horizon tasks and retain reusable experience. The company is combining a proprietary foundation system, structured memory and agentic execution in an effort to move enterprise AI beyond content generation toward persistent workflow automation.
The next challenge for enterprise AI may not be generating better answers. It may be getting AI systems to reliably complete complicated work over hours, across multiple applications, while learning from what happened along the way.
That is the problem Xisiid Intelligence is targeting with a new brain-inspired AI architecture and RMB 25 million seed round. The company, which describes itself as an AI neo-lab focused on new approaches to artificial intelligence, plans to use the funding for research and development of its Brain-Inspired Self-Evolving Foundation System and for joint development of AI applications with domain experts.
Xisiid also recently made its public debut at the Shanghai Pujiang Innovation Forum, where it presented its technical architecture and multidisciplinary research team.
The company’s proposition comes as AI agents move from conversational interfaces toward longer, tool-driven workflows. Rather than treating each task as an isolated interaction with a large language model, Xisiid wants its systems to accumulate experience, retain useful workflows and apply that knowledge to subsequent tasks.
AI Agents Face a Long-Horizon Problem
Current agent systems can perform increasingly sophisticated reasoning, but completing an entire professional workflow remains difficult.
The recently introduced APEX-Agents benchmark illustrates the challenge. Developed to evaluate AI agents on long-horizon, cross-application work in investment banking, management consulting and corporate law, the benchmark requires agents to work with files and software environments rather than simply answer isolated questions. In its initial published evaluation, the best-performing tested agent achieved a Pass@1 score of 24%.
Harvey’s Legal Agent Benchmark points to a similar limitation in legal workflows. Its initial results reported that frontier models completed less than 10% of evaluated legal tasks end-to-end under a strict “all-pass” standard, where every required criterion must be satisfied.
These evaluations matter because professional work often involves more than reasoning about a single prompt. Agents must find relevant information, maintain context, use multiple tools, produce an output and verify that the result satisfies a series of requirements.
Xisiid’s architecture is designed around that problem.
A Dual-System Architecture
The company’s foundation system consists of two principal components, which Xisiid calls System 1 and System 2.
System 1 is designed as a high-speed execution layer. Xisiid says its proprietary latent exchange protocol allows different functional modules to exchange information at the internal representation level rather than repeatedly converting information into natural-language prompts and responses.
That approach is intended to reduce communication overhead in complex agent workflows. Conventional agent architectures frequently coordinate tasks through text-based context, which can become increasingly expensive and difficult to manage as the number of steps grows.
System 2 addresses a different problem: memory.
Rather than storing only factual information, Xisiid says its hierarchical memory layer is designed to retain reusable workflows, decision heuristics and experience from previous task execution. When a new task resembles earlier work, the system can retrieve relevant experience and use it as guidance or constraints.
The two systems are connected through a continual-learning cycle that Xisiid describes as execution, feedback, generalization and consolidation.
The company says evaluation, gating and rollback mechanisms are used to control how new knowledge and capabilities are incorporated. That is significant for enterprise deployment because a self-evolving system cannot simply modify its behavior without mechanisms for validation and recovery.
From Foundation Models to Persistent Intelligence
Xisiid’s approach places memory and continual learning at the center of its AI architecture.
The company argues that increasing model size alone will not solve the reliability problems associated with long-horizon execution. Instead, its architecture attempts to separate rapid task execution from persistent organizational experience.
Xisiid says its initial benchmark validation used models of approximately 30 billion parameters and produced performance it describes as comparable to frontier systems with more than one trillion parameters on selected complex professional tasks. That is a company-reported result, however, and the release does not provide enough independent evaluation methodology to establish a direct apples-to-apples comparison.
The distinction is important as enterprise AI increasingly moves toward specialized systems. APEX-Agents itself was created specifically because conventional model evaluations often do not capture whether an agent can navigate messy, multi-application professional environments and deliver complete work products.
For Xisiid, the opportunity is to build an AI system that becomes more useful as it accumulates experience rather than requiring every new task to begin with a largely stateless model interaction.
Building AI Around Enterprise Memory
The company is already working with institutional clients in areas including investment decision-making, legal due diligence and enterprise operational governance, according to the announcement.
Those applications are particularly relevant to Xisiid’s architecture because they involve knowledge-intensive workflows where organizational context can be as important as general model knowledge.
The company’s on-premise deployment option also targets organizations that want sensitive documents, proprietary workflows and accumulated AI experience to remain within their own infrastructure.
That model could become increasingly relevant as enterprises experiment with AI agents that interact with business-critical information. The technical requirements extend beyond model intelligence to include access controls, auditability, data governance, rollback mechanisms and reliable integration with existing enterprise systems.
Xisiid’s long-term proposition is therefore broader than another AI assistant. It is attempting to create a system in which an organization’s interactions with AI gradually become a reusable knowledge and workflow asset.
Whether that architecture can consistently outperform conventional agent stacks will require independent testing across broader workloads. But the company’s approach reflects an important direction in AI development: moving from models that generate responses toward systems designed to remember, execute, learn and adapt across recurring real-world work.
Market Landscape
Enterprise AI is increasingly being evaluated on task completion rather than isolated response quality. APEX-Agents tests long-horizon professional workflows across investment banking, consulting and law, with agents required to navigate realistic files and applications. Its initial research found that frontier models completed fewer than 25% of the tested tasks, underscoring the difficulty of end-to-end agentic work.
Memory is emerging as another important layer in this transition. AI agents need persistent context, reusable workflows and mechanisms for learning from previous execution if they are expected to handle recurring enterprise processes. Xisiid’s System 1/System 2 architecture is positioned around that combination of execution and structured memory.
The key competitive question will be whether such architectures can deliver measurable gains in reliability, cost and task completion compared with conventional LLM-plus-agent frameworks.
Top Insights
- Xisiid is developing a brain-inspired AI architecture combining fast agent execution with hierarchical memory for continual learning.
- Its RMB 25 million seed round will fund foundation-system development and collaborative AI applications with professional-domain experts.
- APEX-Agents and Harvey LAB demonstrate why reliable completion of long-horizon professional tasks remains difficult for current AI agents.
- Xisiid’s architecture aims to turn validated task experience into reusable workflows rather than treating every AI interaction independently.
- On-premise deployment could allow enterprises to retain proprietary workflows, sensitive information and accumulated organizational AI knowledge internally.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
