OpenClaw Mania and the Missing Memory Layer
Personal AI Agents That Can’t Remember Yesterday
Welcome to another special edition of The Ascent, from AVV. Instead of our monthly look at Vietnam tech, we’re diving into the explosive popularity of OpenClaw in China, the fundamental shortcomings of personal AI agents, and the teams working to add memory to agentic AI.
In case you missed it, check out our article from last week: What LPs Are Missing About Southeast Asia’s Funding Downturn.
By Binh Tran and Mike Tatarski
Over the last two years, AI has evolved from text-based chatbots into systems capable of acting on a user’s behalf. This shift from passive tools to autonomous software has introduced a new tech category: personal AI agents.
One open-source project is now the unlikely focal point for this transition. OpenClaw, an agent framework that allows large language models to plan and execute tasks, has spurred experimentation among developers and companies around the world. Nowhere has its rise been more dramatic than in China.
There, OpenClaw has quickly transformed from a niche developer tool into a nationwide experiment in agentic AI known as “raising lobsters,” after the project’s claw-shaped logo.
Major technology companies such as Tencent, Baidu, Alibaba, MiniMax, and Zhipu have introduced their own agent platforms modeled after OpenClaw, often integrating domestic language models while maintaining the same core idea: autonomous AI agents coordinating tasks across software systems.
Local governments have encouraged adoption, offering subsidies, training programs, and public events aimed at accelerating development.
Enter the Personal AI Agent
Rather than focusing on improving language models themselves, OpenClaw allows developers to build software agents that can act autonomously using those models as reasoning engines.
An OpenClaw agent runs locally on a machine and can perform tasks such as browsing the web, analyzing data, generating reports, interacting with APIs, and executing scripts or workflows. Instead of generating text, these agents plan sequences of actions and execute them across multiple software environments. For developers, this represents a major shift in how AI is used. Rather than assisting with work, AI can increasingly complete the work itself.
Yet the excitement surrounding OpenClaw has also highlighted a technical challenge that affects nearly every AI system: memory.
The Real Bottleneck: Memory
The heart of the problem is that LLMs are stateless by default. Every conversation starts from zero unless you deliberately feed them context. As a result, the gap between a generic AI response and a useful one comes down to what the model knows about you, your work, and the situation before it starts generating.
As these models get more powerful, memory, not intelligence, is the bottleneck. The teams and products that figure out how to feed the right context to the right model at the right time will come out on top.
Despite their impressive capabilities, most personal AI agents today can perform complex tasks during a single session, only to forget what they learned once that ends and start from scratch the next day.
A common workaround is brute force, or dumping everything into the prompt. This creates its own cascade of problems, as models get distracted by irrelevant information, lose track of what actually matters, and start hallucinating connections between unrelated memories.
This is a structural failure that companies like ByteRover, Mem0, and Supermemory are trying to fix.
How Teams Are Building Memory for AI
These teams are building memory layers for personal AI agents, each with a fundamentally different philosophy about how memory should work.
The most established approach is vector embeddings, which store memory in a database. When the agent needs memory, it runs a similarity search to find the closest matches. This works at a basic level, but similarity searches find things that sound alike, not things that are related.
OpenClaw’s default memory takes a simple approach: human-readable files on disk. Your agent writes daily memory logs and curates long-term memory to a special memory file that can be opened and edited in any text editor. However, daily logs accumulate quickly and become unwieldy over time.
Some companies tackle this problem by building cloud-native, API-first systems as a managed service for memory. The tradeoff is clear: you get sophistication and scale, but your memory lives on someone else’s infrastructure.
Then there’s Da Nang-based ByteRover, which takes the most structurally ambitious approach of the bunch. Disclaimer: AVV is an investor in the company.
Like OpenClaw, ByteRover stores memory as local files. But instead of dumping everything into a flat directory of daily memories, ByteRover's agent actively curates and organizes information into what it calls a context tree, a hierarchical structure stored as files in a directory.
When new information comes in, a ByteRover agent reasons about where it belongs in the hierarchy and places it accordingly in a human-readable file. As a16z's Jason Cui and Jennifer Li point out, the most important memories are often implicit, which is why memories belong in human-readable files where people can actually read, refine, and correct them.
ByteRover claims 92.19% retrieval accuracy on the LoCoMo benchmark, and a 99.2% token reduction compared to traditional vector embeddings on a 1,300-file test. If those numbers hold up in broader production environments, they suggest that structure is a fundamental advantage.
Why Structure Beats Similarity
The difference between these approaches boils down to this: should an agent search for relevant memories, or should it know where to look?
A context tree solves this by organizing knowledge the way people actually think, as a hierarchy. As a result, finding the right context becomes navigation rather than search, which is much more precise.
The practical payoff is real: less irrelevant context fed to the AI means lower costs, faster responses, and fewer hallucinated answers. It also means the memory is something a human can actually review and edit to ensure it captures meaning about relationships, not just proximity.
The Next Step: Composable Context Trees
Structured, local memory solves the single-agent problem, but real life isn’t that neat.
Say your company asks you to join an overseas business trip. Your agent needs both your work and your personal contexts, and no single memory store should hold all of that, nor should your private details be open for view
The better model is composable trees: separate, independently maintained knowledge structures that selectively combine based on the task. In this example, both your work and personal trees load together, even though each one has its own owner, update cadence, and access controls.
Security, Governance, and What Remains Unsolved
The rise of personal AI agents has also raised new concerns. Unlike chatbots, agents often require broad permissions to operate effectively, a level of autonomy that introduces obvious risks.
Users have reported incidents where agents accidentally deleted emails or executed unintended actions while performing automated workflows. Governments and corporations have begun discussing safeguards to ensure agents cannot access sensitive data or perform unauthorized tasks.
These issues highlight a fundamental tension in the development of agentic AI: enabling software to act independently while maintaining security and oversight.
The Next Phase of Agentic AI
The rapid rise of OpenClaw demonstrates how quickly new paradigms can reshape the technology landscape. In a matter of months, the framework has sparked a global surge of experimentation with autonomous AI agents.
But the excitement surrounding OpenClaw shows that the agent ecosystem is still in its early stages. For personal AI agents to reach their full potential, they will need new layers of infrastructure, including systems for memory, coordination, security, and governance.
The playing field for companies building those foundational components is competitive. And as the next wave of agent innovation unfolds, the technologies that allow AI systems to remember, learn, and evolve may prove just as important as the intelligence that powers them.
Are you experimenting with OpenClaw? What breakthroughs (or challenges) have you encountered?