Tongyan Lingzhi, a company incubated by the industry platform of the Beijing Institute for General Artificial Intelligence (BIGAI), has laid out its approach to what it calls personal AI: models that keep learning from a specific person, project or company instead of relying only on prompts and retrieval.
Most enterprise AI today adapts to users from the outside. Prompts set rules, retrieval pulls in documents, memory stores state and skills package reusable steps, but the model's own weights rarely change after an interaction. Tongyan Lingzhi argues that stable habits, such as a team's engineering conventions or an investment group's risk discipline, could instead be trained into parameters.
Its architecture keeps a shared base model for general knowledge and reasoning, and adds lightweight Personal LoRA adapters for each individual, project or organization. The company says the aim is for these personal capabilities to survive when the underlying base model is swapped.
To keep those adapters learning, the team is developing recursive self-improvement, or RSI. Task traces, human corrections and outcomes become candidate training data. An evaluator, tuned to each user's or company's standards, filters which signals are worth learning from before an incremental LoRA update. The updated model is rerun on the same tasks as the old version, and only improvements go live, with regressions rolled back. Chief technology officer Chu Bingbing said a single rejected suggestion is only a signal, and that only repeated patterns should become training data.
The company is also working on infrastructure to make thousands of adapters affordable. TIP runs training and inference in parallel on the same GPUs, while Paged LoRA manages adapter memory in pages. In one stage test with Qwen2.5-72B-Instruct on eight Nvidia H800 GPUs and a 5.79GB adapter memory budget, Paged LoRA showed better latency and tail behavior than a flat layout when LoRA requests made up 1% to 30% of traffic, but performance degraded at 50%.
Tongyan Lingzhi is candid that the work is early. It says the RSI technical route and product architecture are set and modules are being engineered, but a fully automated learning loop is still in development, with catastrophic forgetting, feedback contamination and model drift among the open problems.
SEE ALSO: MemTensor Open-Sources MemT 1.0 · Tencent Hunyuan Open-Sources Hy Memory 2.0