The 'Hubble Radius': A New Framework for AI Memory

A Chinese developer and writer known as Pinglunshi has introduced a compelling new framework for thinking about AI agent memory — one that borrows a concept from cosmology to solve a persistent gap in how AI systems understand their users. The concept, called the 'Hubble Radius,' represents a third layer of contextual memory that sits between personal knowledge and public search, capturing the information a user could plausibly have encountered in their daily life.

The idea, published on the WeChat newsletter Hongxian, has sparked discussion in AI developer communities about whether current approaches to personalization and retrieval are fundamentally incomplete.

Key Takeaways

• 3-layer memory architecture: The framework proposes 'what I know,' 'what I should know,' and 'what I might know' as distinct memory tiers for AI agents
• The 'Hubble Radius' is the third layer — it defines the boundary of a user's observable information universe
• Current AI systems typically handle only 2 of these layers: personalization (memory) and search (retrieval)
• The missing middle is awareness of a user's information environment — their feeds, subscriptions, and ambient exposure
• RAG and vector databases solve layer 1, search engines solve layer 2, but no mainstream tool addresses layer 3
• Practical implementation involves feeding an AI agent curated source lists that mirror the user's actual media diet

Why Current AI Memory Falls Short

Most discussions about AI memory revolve around 2 capabilities. The first is personalization — can the AI remember who you are, your preferences, your past conversations? The second is retrieval — can the AI search the internet or a knowledge base to find relevant information?

These map neatly onto existing technologies. Personalization relies on RAG (Retrieval-Augmented Generation) pipelines, vector databases, and persistent conversation histories. Retrieval uses search APIs, web browsing tools, and document stores.

But Pinglunshi argues there is a critical gap between these 2 layers. An AI that only knows what you have explicitly told it is too narrow. An AI that jumps straight to Google when it encounters something unfamiliar is too broad. Neither approach captures a fundamental aspect of human cognition: the ambient information landscape that shapes what you probably already know without ever having explicitly searched for it.

What the Hubble Radius Actually Means

In cosmology, the Hubble radius defines the boundary of the observable universe — the farthest distance from which light can reach an observer. Beyond that boundary, objects exist but are fundamentally invisible to you.

Pinglunshi applies this metaphor to personal information environments. Every person has an 'observable information universe' defined by the sources they regularly consume — their Twitter feed, their RSS subscriptions, the newsletters they read, the Slack channels they monitor, the subreddits they browse.

Information within this radius is something you might plausibly know. You may not have read every article in your feed, but you were exposed to it. An AI that understands this boundary can make smarter inferences about what context you are likely carrying into a conversation — without needing to ask.

This is fundamentally different from both memory and search:

• Memory stores what you explicitly told the AI
• Search retrieves what the entire internet knows
• Hubble Radius captures what your information environment has been broadcasting at you

The 3-Layer Architecture in Practice

The framework breaks down into a concrete implementation pattern that developers can adapt for their own AI agent setups.

Layer 1: 'What I Know' — Personal Knowledge Base

This is the most familiar layer. It includes documents you have written, notes you have saved, conversation histories, and explicit preferences. Technologies like LLM wikis, personal knowledge graphs, and RAG pipelines with vector stores (Pinecone, Weaviate, Chroma) handle this well.

The key principle: this layer should contain information the user has actively created or curated. It represents their explicit, articulated knowledge.

Layer 2: 'What I Should Know' — Structured Reference

This layer covers information the user needs but has not memorized — company documentation, industry standards, technical references, regulatory requirements. It is the domain of traditional RAG retrieval and tool-augmented search.

Unlike Layer 1, this information is not personal. It is shared, canonical, and relatively stable. The AI accesses it on demand rather than holding it in persistent context.

Layer 3: 'What I Might Know' — The Hubble Radius

This is the novel contribution. Layer 3 requires the AI to maintain an awareness of the user's information environment — not the content itself, but the sources and channels through which the user encounters the world.

In practice, this might mean:

• Indexing the user's RSS feeds or newsletter subscriptions
• Monitoring the same Twitter/X lists the user follows
• Tracking headlines from the user's preferred news sources
• Scanning recent posts from the user's professional communities
• Noting trending topics within the user's industry vertical

The AI does not assume the user has read any specific piece of content. Instead, it uses this layer to calibrate its responses — understanding what background context is likely versus unlikely to be familiar.

Why This Matters for AI Agent Design

The Hubble Radius concept addresses a problem that becomes increasingly visible as AI agents take on more complex, autonomous tasks. When an agent operates on behalf of a user, it needs to make judgment calls about what information to surface, what to explain, and what to assume.

Without a sense of the user's information environment, agents face an impossible choice. They either over-explain things the user probably already knows (which feels patronizing and wastes tokens), or they under-explain things the user has never encountered (which leads to confusion and errors).

Consider a concrete example. If a user follows 15 AI researchers on Twitter and subscribes to 3 ML newsletters, their AI agent should probably not spend 200 tokens explaining what 'chain-of-thought prompting' means when it comes up in a task. The concept is almost certainly within the user's Hubble Radius.

Conversely, if the user's information diet is focused entirely on frontend web development, the agent should recognize that advanced ML concepts fall outside their observable universe and provide appropriate context.

This kind of calibration is something good human assistants do instinctively. They develop a mental model of what their boss reads, what conferences they attend, and what professional circles they move in. The Hubble Radius formalizes this intuition into a system architecture.

How This Fits Into the Broader AI Memory Landscape

The concept arrives at a moment when AI memory is becoming a central competitive battleground. OpenAI has added persistent memory to ChatGPT. Google's Gemini offers cross-session personalization. Startups like Mem, Rewind (now Limitless), and Personal AI are building entire products around the premise that AI should know you deeply.

But most of these efforts focus squarely on Layers 1 and 2 — remembering what you have said and retrieving what you need. The Hubble Radius suggests that a third dimension is missing from all of them.

Compared to approaches like Microsoft's Copilot, which integrates with organizational data in Microsoft 365, the Hubble Radius is more personal and more ambient. It is not about documents you have access to — it is about the information weather you live in every day.

Some emerging tools hint at this direction. Artifact (the now-shuttered news app from Instagram's co-founders) tried to model users' information diets. Nuzzle (acquired by Twitter in 2021) surfaced articles shared by people in your network. But neither framed the concept as a memory layer for AI agents.

Practical Implications for Developers and Builders

For developers building AI agents, the Hubble Radius framework suggests several concrete steps:

• Audit your memory architecture: Are you only handling Layers 1 and 2? What would a Layer 3 implementation look like for your use case?
• Build source profiles, not just user profiles: Track not just what users know, but where they learn
• Use Layer 3 for response calibration: Adjust explanation depth based on whether a topic falls inside or outside the user's information radius
• Keep Layer 3 lightweight: This layer should be metadata-heavy and content-light — source lists, topic frequencies, and recency signals rather than full-text indexes
• Respect privacy boundaries: Layer 3 is inherently sensitive — it maps a user's attention patterns, which requires careful consent and data handling

Looking Ahead: The Future of Contextual AI Memory

The Hubble Radius is still a conceptual framework rather than a production-ready standard. But it points toward an important evolution in how we think about AI personalization.

As AI agents become more autonomous — booking meetings, drafting reports, making purchasing decisions — they will need richer models of their users' cognitive contexts. Knowing what someone has explicitly said is not enough. Knowing what the entire internet says is too much. The sweet spot is understanding the information universe that a specific person actually inhabits.

This aligns with broader trends in agentic AI development, where companies like Anthropic, OpenAI, and Google are all racing to build agents that can act reliably on behalf of users over extended periods. Memory architecture will be a key differentiator in that race.

The question is whether the industry will adopt something like the Hubble Radius as a formal architectural pattern, or whether it will emerge organically as agent builders independently discover the same gap. Either way, the insight is clear: AI memory needs more than 2 layers. It needs to understand not just what you know, but what you could know — the boundary of your observable information universe.