APUS AI Open-Sources Reflex Agent, a Personal Agent Driven by Its Own Decision Model

On October 13, APUS AI (Qilin Hesheng Network Technology Co.) announced that it has open-sourced Reflex Agent, its Personal Agent product. The release is published under the MIT license and ships in three forms — a desktop client, a web version and an SDK — so developers can try it, reproduce the setup and extend it. For APUS AI, this is a concrete step in its effort to build a Personal Agent around what it calls Loop Working.

The most interesting design choice is how the Harness is assembled. Conventional agents rely mainly on code and prompts to organize their harness logic. Reflex Agent instead hands core control to Reflex, its own open decision model, which takes responsibility for model routing, skill selection, result review and execution supervision.

That structure produces a loop the company describes as decision, execution, feedback and re-decision. APUS AI calls the working mechanism Loop Working, framing the agent less as a question-answering interface and more as something that keeps acting toward a user's goal over time. It is also the direction that shapes its fit for long-horizon tasks.

Memory is the second pillar. Reflex Agent works with the Continuum memory engine, which accumulates task information through what the project describes as unlimited memory, keeps original content and its associations losslessly, and feeds relevant history into decisions in as close to its original form as possible, reducing the reliance on compressed summaries.

The company is careful about what unlimited context means here. According to the project description, it does not break the input-length limit of a single model call; rather, it aims to free an agent's memory and task cycle from being bounded entirely by one context window, retrieving and invoking history on demand during long-running work.

On performance, APUS AI's lab says Reflex Agent responded up to three times faster than mainstream agents such as Claude Code and Codex in the scenarios it tested. Those figures come from the vendor's own tests and have yet to be independently reproduced.

The open-source artifacts matter as much as the claims: the agent project lives in the GitHub repository APUS-AI-Lab/APUS-Reflex-Agent, while the Reflex decision model is published under the apus-ailab/APUS-Reflex name. With the MIT license, commercial use and modification carry a relatively low barrier, and other teams can inspect how the decision and memory modules are actually implemented.

Read against the wider agent market, the release reflects a shift in competition from single-turn conversational ability toward long-horizon tasks and memory management. Delegating routing, skill selection, review and supervision to a decision model, and letting a separate memory layer carry history, is a pattern several teams are now testing; what makes Reflex Agent notable is that it opens the whole arrangement, so others can study the engineering structure rather than just the speed number.

Two things are worth watching next: whether the community can reproduce the claimed three-times speed advantage, and how the Continuum memory engine behaves on genuinely long tasks. Unlimited memory and lossless input still have to answer the same cost and retrieval-noise questions that face every agent memory system.