Recently, EverMind, incubated by Shanda Group, released Raven Agent, a deep self-evolving agent based on EverOS. Through an original two-way memory mechanism and evolutionary ability to rewrite its own code, EverMind is trying to push AI from a “passive responsive tool” to an “actively evolving digital life.” According to reports, Raven uses the “four-layer bionic architecture” at the bottom of EverOS to divide the original conversation flow into independent memory units, form a “memory scene” through a clustering algorithm, and construct a deep picture including the user's identity, preferences, skills, and work goals. This kind of memory is bidirectional and dynamic; it not only records users' preferences, but also incorporates them into one's own cognitive model. More importantly, you can extract experience from the success or failure of every interaction, and reflect and improve your own response strategies. Test data shows that Raven can use 1/10 of the token consumption of traditional solutions to achieve an accuracy rate that exceeds the full context.

Recently, EverMind, incubated by Shanda Group, released Raven Agent, a deep self-evolving agent based on EverOS. Through an original two-way memory mechanism and evolutionary ability to rewrite its own code, EverMind is trying to push AI from a “passive responsive tool” to an “actively evolving digital life.” According to reports, Raven uses the “four-layer bionic architecture” at the bottom of EverOS to divide the original conversation flow into independent memory units, form a “memory scene” through a clustering algorithm, and construct a deep picture including the user's identity, preferences, skills, and work goals. This kind of memory is bidirectional and dynamic; it not only records users' preferences, but also incorporates them into one's own cognitive model. More importantly, you can extract experience from the success or failure of every interaction, and reflect and improve your own response strategies. Test data shows that Raven can use 1/10 of the token consumption of traditional solutions to achieve an accuracy rate that exceeds the full context.