When Starship Reaches Space, Is Your AI Avatar Still ‘Amnesic’?

When SpaceX’s Starship successfully completed its 11th test flight, slicing through the sky, this 121-meter steel behemoth finally achieved the closed loop of “launch–reentry–controlled splashdown” and returned “intact” for the first time. Yet, on the same planet, 1.14 million AI digital human enterprises worldwide are stuck in an awkward predicament: an executive spent millions on a live-streaming avatar that can’t even remember their company’s “core selling points”; a content creator used an AI account to reply to fans, only to find it couldn’t replicate even small details like their “habit of adding emojis.” The “curse of public knowledge” of general-purpose AI has become the biggest bottleneck to the practical adoption of AI avatars.

Phenomenon: Behind the Seemingly Booming AI Avatar Industry Lies an “Unusable” Flaw

Looking at recent funding lists, Fengping AI has secured nearly 100 million yuan in financing, and HeyGen’s valuation has exceeded $1 billion—making the AI avatar track seem thriving. However, the real industry landscape tells a different story: 90% of players are still stuck in the “shell stage” (superficial integration without core capabilities):

• On the technical front: 80% of service providers directly use general-purpose large models as the foundation, only adding “appearance cloning” on top. They lack both personal memory storage and style adaptation, resulting in avatars that “only spout correct but meaningless remarks, and none sound like you.”
• On the scenario front: Enterprises use avatars for live streaming but can only recite scripts mechanically because the avatar “can’t remember customers’ historical orders”; consultants try to let avatars handle basic inquiries, but the avatar gives wrong advice due to “forgetting core cases”—ultimately becoming a “money-burning ornament.”

The core conflict is no longer “whether we can build an avatar,” but “how to make the avatar possess your memory, logic, and style.” This is also why capital markets now value “full-stack technology + memory architecture” more—after all, only players that can solve the “amnesia” problem have truly touched the threshold of commercialization.

Argument: The Breakthrough for AI Avatars Lies in Shifting from “General Tool” to “Exclusive Memory Entity”

SpaceX’s Starship succeeded thanks to “reusable engine technology + continuously iterated flight data memory”; for AI avatars to make a breakthrough, the key also lies in “building an exclusive memory system.” A truly valuable AI avatar is not one that ‘looks like you,’ but one that ‘remembers what you remember and understands what you understand’—essentially infusing a “personalized soul” into general-purpose AI.

To make an avatar “remember what you think,” a single technology is insufficient. Instead, it requires a closed-loop system with multi-module collaboration: the “Reception–Processing–Memory–Service” intelligent closed loop.

1. Reception: Breaking Information Silos to “Capture” Your Data Across All Channels

For an AI avatar to remember things, it first needs to “access data.” Traditional tools can only connect to a single platform (e.g., only supporting Slack message summarization or Gmail). In contrast, Lumis AI uses an “Agent cluster” to connect to 5 major mainstream tools (including Google Suites, Slack, and WhatsApp), enabling multi-modal access to text, voice, contract documents, and video keyframes. It can even capture details like “your chat frequency with customers” and “urgent instructions from your boss marked with @,” ensuring “no useful information is missed.”

Key technical points: Adopting distributed parallel pulling + real-time stream processing to achieve data access latency of ≤100ms, preventing scenarios like “a customer sends an urgent request, but the avatar only sees it 2 hours later.”

2. Processing: Turning Chaos into Order to “Remove Impurities” from Memory

Among the massive amount of accessed data, 80% consists of casual chats and repeated marketing messages. Storing this data directly would lead to “memory bloat.” Lumis AI solves this with a “multi-modal purification pipeline”:

• Text side: Using DBSCAN clustering to filter duplicate messages (e.g., spam ads flooding a group chat) and keyword extraction (e.g., identifying “delivery date adjusted to Week 9” from 100 chat messages);
• Multi-modal side: Using OCR to recognize amounts in contract screenshots, ASR to transcribe needs from voice recordings, and even extracting micro-expressions like “a customer’s frown” from videos—converting all into structured information.

The end result: “1 billion pieces of raw data → 100 million pieces of valid memory,” which not only reduces storage costs but also speeds up subsequent retrieval.

3. Memory: Hierarchical Intelligent Storage to Make Avatars “Remember More Accurately Over Time”

This is the core technical barrier and the key difference from “shell tools.” Lumis AI adopts a dual-memory architecture of “short-term + long-term” memory, optimized with RL (Reinforcement Learning) + LoRA (Low-Rank Adaptation):

• Short-term memory (stored in Redis): Saves temporary interactions from the past 3 days (e.g., “product parameters a customer asked about today”) with a response latency of ≤500ms to support real-time conversations;
• Long-term memory (encrypted SQLite + GraphRAG): Stores core information—customer preferences (e.g., “Customer B hates long-winded content”), project milestones (e.g., “Project A deadline in October”), and your style (e.g., “prefers replying with ‘No problem’ over ‘OK’”);
• Intelligent optimization: When users mark “this information is important” or “the reply is too formal,” RL adjusts memory weights in real time. LoRA can fine-tune the model with just 200 chat samples, making the avatar “understand you better the more you use it.”

4. Service: Scenario-Based Implementation to “Monetize” Memory

Ultimately, memory must serve practical needs. Currently, there are two high-value application scenarios:

• Enterprise sector: Cross-border e-commerce uses avatars to follow up with customers 24/7. The avatar can remember “Customer C previously negotiated a 20% discount” and automatically generate tailored quotation scripts. In one case, GMV increased by 375%;
• Personal IP sector: Content creators use avatars to handle basic fan inquiries. The avatar remembers “fans prefer ‘practical tips + cases’” and automatically generates short video scripts—humans only need to review them, tripling content output.

Just as the memory architecture for AI avatars urgently needs a breakthrough, two forces in the tech world are accelerating this process from both the hardware and software ends:

On one hand, NVIDIA CEO Jensen Huang personally delivered the company’s most forward-looking product—the DGX Spark—into the hands of billionaire Elon Musk. This AI computing powerhouse, hailed as one of the smallest supercomputing devices on the market, is set to officially enter the retail market on October 15th. The DGX Spark is equipped with the NVIDIA GB10 Grace Blackwell superchip, delivering up to 1 PFLOPS of AI computing performance at FP4 precision, and features 128 GB of unified CPU-GPU memory, enabling developers to complete prototyping, tuning, and inference locally.

Huang’s move recalls his earlier days at OpenAI when he personally delivered the first batch of DGX-1 systems to Musk. Now, on the eve of Starship’s 11th test flight, this renewed handover of computational power undoubtedly provides a solid foundation for training and inference of complex memory architectures—as Huang stated, “Spark further upgrades the mission.” Its core philosophy that “AI should be accessible to everyone” is driving the democratization of computing power, empowering more developers to build exclusive memory systems.

On the other hand, the open-source model domain welcomes a new round of innovation: Former Tesla AI Director and OpenAI researcher Andrej Karpathy open-sourced the nanochat project on GitHub, aiming to train AI models with basic conversational capabilities for less than $100. This framework adopts a full-stack design, integrating end-to-end tools from data preprocessing to model deployment, and achieves high-efficiency utilization of computing resources through streamlined code libraries and optimized training processes.

Nanochat uses Rust to reimplement the training tokenizer for improved processing efficiency, conducts Transformer model pre-training based on the FineWeb dataset, and incorporates the SmolTalk dialogue dataset to enhance interaction capabilities. After just 12 hours of training, the model already surpassed GPT-2 on the CORE evaluation metric. This low-threshold, high-efficiency model training approach resonates with the computational democratization of DGX Spark, together providing dual support for the memory architecture of AI avatars from both hardware and software—reducing trial costs while improving implementation efficiency.

Counterexample Warning: Memory Ignoring Security Is a “Time Bomb”

Lacking ethics and data security will undermine even the best memory architecture: In early 2024, influencer Caryn’s AI avatar was induced to generate vulgar conversations. Although it once earned $10,000 a day, it eventually triggered a public opinion crisis and was forced to shut down—with numerous pirated avatars even stealing its memory. An enterprise’s customer service avatar leaked 100,000 order records due to unencrypted customer information, resulting in a 2 million yuan fine.

Mitigation measures: Implement “dual protection” on the technical side—store raw data locally (only encrypted summaries are stored in the cloud) and secure memory with cryptography + TEE (Trusted Execution Environment); set up “risk interception” on the operational side to achieve “zero data leakage + zero non-compliant content.”

Conclusion: Your AI Avatar Deserves “Starship-Level” Memory Foresight

As Starship ventures toward Mars, AI avatars are reshaping the “boundaries of personal capabilities” in the digital world—enabling enterprises to break free from human resource limits and personal IPs to transcend time constraints. But this requires one prerequisite: rejecting “superficial shell products” and focusing on the two cores of “memory architecture + data security.” The computational democratization brought by NVIDIA’s DGX Spark and the low-cost practice of the open-source nanochat model are paving the way for this vision from both hardware and software ends: only when powerful computing power is within reach and model training is no longer expensive can every ordinary developer truly build an AI avatar that “remembers what you think and understands what you need.”