The One Gap Everyone Agrees On: Continuous Learning and the Future of AGI Beyond Model Weights

In a rare investor briefing, DeepSeek’s Liang Wenfeng articulated a consensus emerging from disparate corners of the artificial intelligence research community: the singular missing piece on the path to Artificial General Intelligence (AGI) is not a larger model, advanced video processing, or another new modality, but rather the ability for models to learn continuously. This assertion marks a significant pivot in the discourse, moving the focus from sheer computational power and data scale to the architectural challenge of dynamic, persistent knowledge acquisition. Wenfeng posited that solving this fundamental limitation would pave the way for a "gradual singularity" and, ultimately, embodied intelligence, where AI systems can integrate and act within the physical world.

This declaration resonates deeply across the AI landscape, finding echoes in the work of prominent researchers and leading organizations. Figures like Andrej Karpathy have conceptualized agents capable of installing new skills akin to .md files, suggesting an externalized, modular approach to knowledge. OpenAI, a pioneer in large language models, has explored similar concepts through initiatives like "Dreaming," a mechanism for background memory synthesis between user sessions. Microsoft, at its annual Build conference, demonstrated efforts to embed agent identity directly into the operating system, implying a persistent, context-aware digital persona. Even open-source projects, such as openclaw, have adopted features like /dreaming, further solidifying the convergence on continuous learning as the frontier’s most pressing challenge. This widespread agreement signals a paradigm shift, indicating that the next major leap in AI will hinge not on model size alone, but on how these intelligent systems accrue and retain experience over time.

The Inherent Limitations of Model Weights for Continuous Learning

Despite the undeniable advancements in large language models (LLMs), a critical insight now firmly held by leading experts is that the "frozen model" – the static, pre-trained neural network – cannot be the locus for continuous learning. This conclusion stems from three insurmountable obstacles: economic impracticality, inherent opacity, and the problem of vendor lock-in. Understanding these limitations is crucial for appreciating why the search for AGI demands a fundamental re-evaluation of how AI systems acquire and manage knowledge.

Firstly, economic impracticality presents a formidable barrier. The training of a state-of-the-art foundation model requires immense computational resources, often involving thousands of high-end GPUs and consuming millions of dollars over several months. This scale of investment makes it utterly unfeasible to retrain a model for every individual user or even for every new interaction session. The core appeal of a foundation model lies in its shared utility, serving a vast user base from a single, pre-computed knowledge base. The moment an AI is expected to remember the nuances of a specific user’s last conversation or a particular enterprise’s evolving operational procedures, the traditional training paradigm breaks down. Fine-tuning offers a limited workaround, but it still represents a batch-based, costly process that doesn’t permit real-time, continuous adaptation. The sheer scale of parameters in modern LLMs, often in the hundreds of billions or even trillions, makes dynamic, on-the-fly weight modification per interaction a computational nightmare, pushing the boundaries of current hardware and energy consumption capabilities far beyond practical limits.

Secondly, the opacity of model weights poses a significant challenge, particularly for applications requiring trust, auditability, and precise control over learned information. Neural network weights exist as an intricate, high-dimensional mathematical representation of patterns and correlations derived from vast datasets. They form a "black box" where individual pieces of learned information cannot be easily isolated, inspected, or modified. One cannot simply "open up" the weights and identify what the model learned about a specific topic last Tuesday, let alone delete a single incorrect fact or audit what it retained about a user’s sensitive data. For AI systems intended to function as reliable coworkers, customer support agents, or components of critical systems of record, this lack of transparency is disqualifying. Regulatory frameworks, such as the European Union’s AI Act or data privacy regulations like GDPR, increasingly demand explainability and the "right to be forgotten," which are fundamentally incompatible with the monolithic, uninspectable nature of trained weights.

Finally, vendor lock-in emerges as a strategic risk. If an agent’s accumulated experience and learned skills are inextricably bound within the specific parameters of one vendor’s model, that knowledge becomes fragile and transient. The moment an organization decides to switch to a different, potentially superior, or more cost-effective model – a common necessity in the rapidly evolving AI landscape – all that accumulated experience is effectively lost. The field of AI moves with breathtaking speed; new models are released, architectures are optimized, and performance benchmarks are shattered on a near-monthly basis. Marrying an agent’s entire operational memory to a single, static set of weights is a recipe for obsolescence, preventing agility and competitive adaptation. In essence, continuous learning that is confined to the weights of a foundation model is, from the user’s perspective, tantamount to continuous forgetting, as every new session or model iteration risks starting from square one.

The File Layer: A Durable, Inspectable, and Portable Solution

Given the inherent limitations of embedding continuous learning within model weights, the consensus solution points towards externalizing this memory to a "file layer." This approach advocates for a store that is durable, inspectable, portable, and fundamentally model-agnostic, allowing knowledge to persist independently of the specific AI engine processing it. Increasingly, this store takes the form of plain files, leveraging established principles of data management.

A compelling example of this emerging pattern is Tolaria, an open-source knowledge base that recently garnered attention. Tolaria’s design embodies the core tenets of this files-first philosophy: it utilizes markdown files for content, a Git repository for version control, YAML frontmatter for metadata, an MCP server for structured interaction, and AGENTS files to facilitate reading and writing by external AI tools. Its foundational design principles are articulated like a manifesto: "files-first, git-first, offline-first, zero lock-in." This means no proprietary accounts, no mandatory cloud dependencies, and complete ownership of one’s knowledge, stored simply as files on a local system or a self-managed repository.

This architectural choice is far from coincidental; it represents a logical progression in the "files-first" movement within software development and now, increasingly, AI. Files inherently solve the three critical problems that model weights cannot:

1. Portability: Files are universally accessible and editable by any text editor, integrated development environment, or, crucially, any AI model. This eliminates vendor lock-in, allowing an agent’s accumulated knowledge to seamlessly transition between different underlying LLMs without loss of fidelity or functionality.
2. Inspectability: Unlike opaque model weights, files are human-readable. Every piece of information, every learned fact, and every modification can be directly accessed, audited, and diffed (compared) by a human or an automated system. This transparency is vital for debugging, ensuring accuracy, maintaining compliance, and building user trust.
3. Version Control: The integration with Git, a distributed version control system, provides a robust, battle-tested solution for managing memory over time. Git inherently handles changes, reverts, branching, and collaboration, effectively providing a granular, auditable history of an agent’s evolving knowledge base – a feature that has been central to software development for decades.

This approach has been rigorously tested in practical applications. For instance, an agent runtime was rebased across four months of upstream changes, and its underlying inference model was entirely swapped out. Despite significant alterations to both the operational environment and the core AI engine, the agent’s memory, stored in plain markdown files and Git-synced, remained perfectly intact. Crucially, the agent continued to cite information it had learned weeks earlier, under a completely different model. This demonstration underscores the fundamental principle: for true continuous learning, the learning must survive the model itself.

Beyond Storage: The Cognition Layer – "Soul Memory"

While the move to a file layer addresses the fundamental issues of durability, inspectability, and portability, it is crucial to recognize that merely storing information in files does not equate to continuous learning. A folder full of notes, however well-organized, is essentially a sophisticated filing cabinet, not a cognitive system. The real, challenging, and valuable work lies in the runtime layer that orchestrates the interaction with these files – what some refer to as "Soul Memory."

This cognition layer is what transforms static data into dynamic, actionable intelligence. It encompasses a suite of sophisticated processes that go far beyond simple retrieval:

• Contextual Retrieval and Filtering: Intelligently identifying and extracting the most relevant pieces of information from a vast knowledge base based on the current context, task, and user query, rather than a brute-force search.
• Information Synthesis and Reasoning: Connecting disparate facts, identifying patterns, drawing inferences, and generating novel insights from the stored knowledge. This involves moving beyond mere recall to genuine understanding and problem-solving.
• Knowledge Prioritization and Forgetting Mechanisms: Deciding what information is most critical, what can be summarized, and what can be safely archived or forgotten to prevent cognitive overload and maintain efficiency. This also includes mechanisms for proactively identifying and correcting outdated or erroneous information.
• Self-Correction and Refinement: Continuously evaluating the utility and accuracy of stored memories, updating them based on new experiences, feedback, or logical inconsistencies.
• Integration with External Tools and APIs: Enabling the agent to not only read from its memory but also to actively use that memory to interact with external systems, perform actions, and gather new information.
• Managing Memory Hierarchies: Implementing different types of memory, such as short-term (episodic), long-term (semantic), and procedural memory, to mimic human-like cognitive structures.

The markdown store itself is a commodity – anyone can write files. The true innovation and competitive moat reside in this sophisticated cognition layer, which imbues the agent with the ability to truly "learn" and "remember" in a meaningful way. It’s the difference between a library and a scholar.

The Counterintuitive Part: Bad Memory is Worse Than None

A critical insight derived from recent experiential memory research is that not all memory is beneficial; in fact, poorly managed memory can be detrimental. In a blind study, agents provided with "synthetic memory" – typically tidy, summarized recaps of past interactions or learned facts – performed demonstrably worse than agents given no memory at all. The conclusion was striking: a confident, yet lossy, summary proved more damaging than an honest blank slate. This is because the compressed summaries often strip away crucial caveats, contextual dependencies, and indicators of uncertainty, leading the agent to become overconfident about information it had subtly misunderstood or whose applicability was limited.

This represents a significant trap in naive approaches to continuous learning. When complex experiences are distilled into neat summaries, the first elements to be discarded are often the nuanced qualifiers: "this was uncertain," "this only held in that specific context," or "I didn’t actually know this part definitively." However, when the agent later retrieves and reads these summaries, it interprets the surviving, simplified sentences as undisputed facts. This process does not make the agent smarter; instead, it fosters a dangerous overconfidence in precisely the areas where its understanding was flawed.

Therefore, how an AI remembers is as crucial as what it remembers. An effective memory format must preserve, alongside each fact, its provenance (where it originated and the degree of certainty), its scope (the specific conditions or contexts under which it applies), and an explicit mechanism to express uncertainty or the lack of specific information ("I don’t have that"). These elements are not mere decorative additions; they are fundamental safeguards that prevent a concise summary from solidifying into false confidence, enabling the agent to maintain a realistic assessment of its own knowledge and capabilities.

Implications for Developers Building the Next Generation of Agents

The overarching message for developers and organizations building AI agents is clear and actionable:

1. Own Your Agent’s Memory: Do not allow critical learned experience to be locked within proprietary model weights. Design systems where the agent’s knowledge base is independent and under your control.
2. Prioritize Portability and Auditability: Ensure that the agent’s memory is stored in a format that is universally readable, inspectable, and easily transferrable across different models and platforms. File-based, Git-versioned systems are ideal for this.
3. Invest in the Cognition Layer (Runtime): Recognize that the true value lies not in the raw storage of files, but in the sophisticated software layer that enables intelligent retrieval, synthesis, self-correction, and contextual application of that memory. This "Soul Memory" runtime is the critical differentiator.
4. Emphasize Memory Quality Over Quantity: Design memory systems that preserve context, provenance, and uncertainty. A smaller, well-structured, and accurate memory is far more valuable than a vast, but flawed, collection of summarized facts.

Continuous learning is no longer a distant aspiration but the acknowledged gap that the entire AI field is now converging upon. However, it is not a problem awaiting the next, even larger foundation model or a breakthrough in neural network training. Instead, it is fundamentally a systems problem, rooted in the architectural decisions made at the file layer, and critically, in the intelligence of the memory runtime. The teams that will lead the charge towards AGI will not necessarily be those with the biggest weights or the largest training datasets. They will be the ones whose agents remember effectively, remember well, and retain that accumulated wisdom across every model they ever utilize, enabling true, durable intelligence to flourish. This shift represents a profound reorientation, moving beyond static models to dynamic, evolving, and truly intelligent agents.