According to BlockBeats, on August 30th, GMGN monitoring showed that crypto KOL Unipcs (@theunipcs) currently has a floating profit of approximately $3.403 million on his PONS holdings, representing a return of about 52 times. He previously spent 67,304 USDG to buy 10.96 million PONS and has not sold any tokens to date; this holding is currently worth approximately $3.47 million. His account asset value has exceeded $7.3 million, a new all-time high.

Unipcs gained fame in 2025 with the hype surrounding Solana's ecosystem meme coin USELESS, at which time it saw a floating profit of $7.5 million on a single USELESS token.

According to Beating AI News, Ashwin Gopinath, a former MIT professor and co-author of Reflexion, has proposed a hypothesis: What truly sets Instinct, the personal agent assistant that has recently become a sensation in Silicon Valley, apart from other agents may not be its execution capabilities or ordinary long-term memory, but rather its continuous maintenance of the user's "long-term state".

Gopinath was founder Noah Shinn's mentor, and the two collaborated on Reflexion research in 2023. He lacks internal information about Instinct; the following assessment is based on actual user experience:

Looking solely at task execution, Instinct doesn't offer any particularly remarkable capabilities compared to agents like Hermes and OpenClaw. Many agents now also perform replays, save memories, and even automatically generate skills. What makes Instinct truly unique is that the longer it's used, the better it seems to understand the user, without being noticeably burdened by ever-increasing amounts of historical information.

The key may be that it doesn't simply save all chat and operation records and then retrieve them using vector search or knowledge graphs. Instead, it continuously maintains a "current state": which user preferences are still valid, which facts have changed, which tasks have not yet been completed, and whether the next step should be to wait, continue to follow up, or take direct action.

This is crucial for long-term agents. A task that lasts several days or even weeks may involve email replies, waiting, changing conditions, and multiple follow-ups. The agent not only needs to remember what happened in the past, but also needs to know which information is outdated, what is truly important now, and when it is absolutely unnecessary to bother the user.

Gopinath calls this process "memory compilation": constantly sifting useful information from a jumble of life records and compressing it into a continuously updated state. The focus is not on "remembering more," but on knowing what is worth keeping, what should become obsolete, and what changes should trigger action.

This is a clear continuation of Reflexion from three years ago. Reflexion allowed agents to summarize failures into written memories for future attempts; with long-term personal agents, the problem has further evolved into how to continuously understand a constantly changing person and the real world.

As large-scale model capabilities become increasingly similar, and tool calls and execution frameworks are rapidly becoming more widespread, the truly difficult-to-replicate aspect may lie in who can consistently remember the right things and take action at the right time.