On November 2,
The service extracts key information from conversations in real time and automatically saves it as user profiles and structured memories according to rules. In subsequent interactions, the system can accurately retrieve and use relevant context based on the user's identity and current question, enabling continuous follow-up on user needs and effectively reducing repeated questions.
Evaluation data shows that in the recently released [HaluMem-medium] comparison, Memory achieved the highest scores in recall, accuracy [F1] metrics for memory extraction, and the memory QA C metric; in the PersonaMem-v2] evaluation, it ranked first among the [8] evaluated systems with an overall score of 56.82], and took first place in [7] of the 9 task dimensions.
In real-world application scenarios, taking financial advisory as an example, the system can automatically summarize a user's risk preferences and liquidity needs and generate a profile. In subsequent conversations, it works with an [Agent] to complete product comparisons and risk reminders, and dynamically updates memories when the user's preferences change. Currently, developers can not only customize memory rules, profile fields, and track retrieval performance in the console, but also seamlessly integrate this capability via API, MCP, or CLI into mainstream [Agent] tools and frameworks such as Codex, DeepSeek Harness, Claude Code], and OpenClaw]. In addition, Memory] can work closely with the Skill] module to decouple individual preferences from reusable task workflows and enable efficient collaboration.
This upgrade marks that personal assistants are accelerating past the context bottleneck of traditional single-turn conversations, evolving toward both deep personalization and long-term companionship capabilities, and providing solid infrastructure support for [Agent]s to go deep into complex vertical scenarios.
