Summary: Current AI systems possess superhuman memory in two forms, parametric knowledge from training and context windows holding hundreds of pages, yet no pathway connects them. Everything learned in-context vanishes when the conversation ends, a computational form of anterograde amnesia. Recent research suggests weight-based continual learning may be closer than commonly assumed. If these techniques scale, and no other major obstacle emerges, the path to AGI may be shorter than expected, with serious implications for timelines and for technical alignment research that assumes frozen weights.

Intro

Ask researchers what's missing on the path to AGI, and continual learning frequently tops the list. It is the first reason Dwarkesh Patel gave for having longer AGI timelines than many at frontier labs. The ability to learn from experience, to accumulate knowledge over time, is how humans are able to perform virtually all their intellectual feats, and yet current AI systems, for all their impressive capabilities, simply cannot do it.

The Paradox of AI Memory: Superhuman Memory, Twice Over

What makes this puzzling is that large language models already possess memory capabilities far beyond human reach, in two distinct ways.

First, parametric memory: the knowledge encoded in billions of weights during training. Leading models have ingested essentially the entire public internet, plus vast libraries of books, code, and scientific literature. On GPQA Diamond, a benchmark of graduate-level science questions where PhD domain experts score around 70%, frontier models now exceed 90%.

They write working code in dozens of programming languages and top competitive programming leaderboards. They are proficient in most human spoken languages and could beat any Jeopardy champion. Last year, an AI system achieved gold-medal performance at the International Mathematical Olympiad, and models are increasingly reported to be helpful when applied in cutting-edge math and physics research. A frontier model's parametric memory contains more facts, patterns, and skills than anyone could acquire in several lifetimes.

Then there's working memory: the context window. Here too, models are superhuman. A 128,000-token context holds the equivalent of a 300-page book in perfect recall, every word accessible, every detail retrievable. Within that window, models exhibit remarkable in-context learning: a few examples of a new pattern and they generalize it, often with human comparable efficiency. A handful of demonstrations can teach a new writing style, notation, codebase convention or abstract pattern.

So what gives? The answer is that these two memory systems are entirely disconnected. There is no pathway from context to weights. When the conversation ends, everything in the context window vanishes.

The closest analogy is anterograde amnesia, the condition where patients can access old memories and function normally moment-to-moment, but cannot form new ones. They can hold a conversation, follow complex reasoning, even learn within a session. But the next day, it's gone. Every morning begins from the same fixed point. Current AI systems suffer from an extreme^[1]computational version of exactly this.

The primary workaround is externalized memory. Patients with anterograde amnesia do not regain the ability to form new memories. Instead, they offload memory into persistent artifacts, diaries, notebooks, whiteboards, written logs, that can be reread to reconstruct context after each reset. These aids function as a surrogate episodic store: nothing is learned internally, but enough state can be recovered to function.

This is essentially what we've built for AI systems.

The Scaffolding Approach

The industry has converged on increasingly sophisticated scaffolding—external systems that manage context on the model's behalf. Andrej Karpathy has described this as treating LLMs like a new operating system: the model is the CPU, the context window is RAM, and "context engineering" is the art of curating exactly what information should occupy that precious working memory at any moment.

These scaffolds take several forms. Retrieval-Augmented Generation (RAG) stores information externally and injects relevant chunks into prompts on demand. Memory features (now standard in Claude, ChatGPT…) summarize conversations into persistent user profiles, inserted in the context window at the start of each conversation. CLAUDE.md/AGENTS.md files act as onboarding documents for coding agents, loaded anew each session. Agent Skills package instructions and resources into discoverable folders for progressive disclosure, surfacing relevant context on demand rather than loading everything upfront. And when agents hit context limits, compaction aggressively summarizes trajectories, compressing hours of work into a few thousand tokens while preserving key decisions, file changes, and pending tasks.

What emerges is a complex system where the static model sits at the center of a web of external memory aids. Context flows in from files, databases, and retrieval systems; gets processed; then flows back out to storage. These systems genuinely work, and alongside their superhuman parametric knowledge and context window, allow them to serve as increasingly competent coding agents that can now maintain coherence on knowledge tasks far better than any human with anterograde amnesia could.

The frontier of this approach is agentic context management: rather than building scaffolding that manages memory for the model, train the model to use tools to manage its own context. Context-Folding, for instance, trains agents to branch into sub-trajectories for subtasks and then collapse the intermediate steps into concise summaries upon completion, effectively needing 10× smaller active context. Instead of external compression imposed by the system, the model actively decides what to keep, delegate, or discard.

But all of this, however sophisticated, remains fundamentally external. The model itself stays frozen. Every capability for persistence lives outside the weights.

Is This Enough?

There's a plausible argument that these scaffolds, combined with scaling and better training, will suffice for any practical purpose. We didn't need to figure out how to fly by flapping wings to build airplanes; we needed powerful engines and large fixed surfaces for lift.

Perhaps we don't need weight-based continual learning at all and sufficiently large context windows, sufficiently good retrieval, sufficiently intelligent context management, all wrapped in increasingly capable base models, gets us wherever we need to go. The 1M+ token windows now emerging make RAG optional for many use cases. Agentic context folding can extend effective memory arbitrarily far.

But there are also reasons to think scaffolding alone won't close the gap. Context windows remain bounded, by compute, memory, and the difficulty of retrieving the right information from millions of tokens. And while in-context learning is remarkably flexible, it likely operates within the space of representations the model already has, recombining and adapting rather than building fundamentally new cognitive machinery. Weight-based learning could offer something qualitatively different: the ability to reshape the model's computational structure itself, forming new abstractions that didn't exist before. The prize would be algorithms that bridge these modes, retaining the sample efficiency of in-context learning while allowing experience to gradually restructure the system.

Weight based continual learning

No frontier model currently has any such weight-based learning capability. However, recent papers have begun exploring ways to give LLMs more continuous learning, leveraging the enormous parametric capacity of deep networks to store memories directly in weights. Here, I want to highlight two papers, the second building on the first, which in my shallow review seem especially promising and whose techniques I expect to appear in frontier models in the not too distant future.

Titans

This Google Research paper hit arXiv on the last day of 2024. Titans proposes a hybrid architecture with a "Neural Long-term Memory" module that updates its own parameters during inference. When the model encounters surprising input (operationally, inputs that induce a large learning signal under an auxiliary associative (key→value) objective, where the key/value projections and the update/forgetting dynamics are meta-trained in the outer loop), it performs a small gradient-based update to store that association in the memory network’s weights. Retrieval is then just a forward pass through this memory network.

While most linear recurrent models (e.g., linear attention/SSMs) compress history into a fixed vector- or matrix-valued state via a predefined recurrence, Titans instead uses a deep neural memory whose parameters themselves serve as the long-term state. This yields a two-level optimization view: in the outer loop (standard training), the full model is trained end-to-end, but it is also trained so that the memory module becomes a good online associative learner. Then at inference time, the memory module updates its own weights using small gradient steps on that auxiliary key–value reconstruction loss (rather than the LM loss), with momentum and adaptive decay acting as memory persistence/forgetting. In effect, the model learns a learning procedure for what to store and what to forget in long-term memory.

In the "Memory as Context" (MAC) variant, this long-term storage is integrated with a traditional transformer attention block over segment-sized windows. Before processing a new segment, the model uses the current input as a query to retrieve relevant historical abstractions from the neural memory. These retrieved embeddings are prepended to the input sequence as pseudo-tokens, effectively providing the short-term attention mechanism with a summarized "historical context." This allows the model to maintain the high-precision dependency modeling of attention while grounding it in a much deeper, non-linear representation of the distant past.

On typical language modeling and commonsense reasoning benchmarks, the Titans variants are generally competitive with strong Transformer and modern recurrent baselines. The bigger separation shows up on long-context retrieval and reasoning benchmarks: on RULER’s needle-in-a-haystack tasks, Titans, especially the MAC/MAG variants, keeps high retrieval accuracy out to 16K tokens in settings where several linear-recurrent baselines degrade sharply with length. In BABILong, a harder “reasoning-in-a-haystack” benchmark where the model must combine facts scattered across very long contexts, Titans (MAC) is reported to outperform a range of baselines in both few-shot and fine-tuning regimes. In the fine-tuned results, at contexts of 1M tokens, accuracy remains above 90%. The authors also claim the approach can be effectively scaled to over 2M tokens.

In terms of compute, they show that while Titans (MAC) is slightly slower to train per-token than the fastest linear recurrent models (like Mamba2 and Gated DeltaNet) due to the complexity of updating a deep network, it avoids the sequence length penalty of Transformers. The architecture decouples the operations: high-fidelity attention runs only on local segments, while the "write" to long-term memory is amortized over chunks. This allows for a still competitive throughput while yielding an increasing advantage as context increases. While a standard Transformer’s KV cache grows linearly, the Titans "state" is just the fixed-size weights of the memory MLP. The compute advantage can become very large for long contexts, even with modern more efficient self-attention variants, as prefill scales linearly rather than quadratically, and generation avoids the per-token bandwidth cost of reading an ever-growing KV cache.

However, Titans’ “test-time memorization” introduces costs that standard inference avoids. While they introduce hardware-friendly optimizations, inference still includes a backward-style computation through the memory module, which involves a momentum-like surprise accumulator that is the same size as the memory parameters. And at the infrastructure level per-user or per-session memory states would break the efficiency gained from serving batches for many users together through shared weights, requiring either separate computation or some new optimization techniques.

The largest experiment used a 760M parameter model trained on 30B tokens. I haven't found anyone applying this at larger compute budgets yet, but Jeff Dean mentions the paper as something they could incorporate into future Gemini models, so we might see something like it in a frontier model before long.

Nested Learning / Hope

Published by the same authors as Titans, Nested Learning is a more ambitious conceptual paper that challenges the binary distinction between short-term (context) and long-term (weight) memory. It proposes a unified framework suggesting that what we call "architecture" (and even “optimizer”) is better understood as a collection of nested optimization problems running at different update frequencies. I drew heavily on their neurophysiological motivations to write the opening of this post.

To understand their proposed solution, we can first look at how they reframe the standard Transformer. From the Nested Learning perspective, a standard Transformer is defined by two disparate levels of optimization with extreme update frequencies: infinite and zero. The Attention mechanism represents the infinite-frequency extreme because it functions as a non-parametric associative memory that instantaneously "solves" a regression objective to map keys to values for every new token, adapting perfectly to the immediate context but retaining no persistent state once that context is removed.

Meanwhile, the MLP blocks and attention weight projection matrices represent the zero-frequency extreme during inference because their parameters are optimized only during the outer "pre-training" loop and remain frozen at test time, representing static long-term parametric memory that cannot be modified by new experiences. This design creates a gap where the model exhibits "anterograde amnesia." It possesses fleeting short-term adaptation and fixed long-term knowledge, but lacks the intermediate, parametric updates necessary for continual learning or memory consolidation.

To bridge this gap, the authors propose Hope, an architecture that fills the "frequency void" between attention's fleeting adaptations and the frozen permanence of pre-trained weights. While the Transformer operates at two extremes, they design Hope to work with a continuum of memory plasticity where different parts of the model are trained at different rates.

The architecture achieves this by replacing the two standard components of the Transformer with dynamic counterparts:

Self‑Referential Titans (Replacing Attention in Hope): Standard Titans adds an online neural memory that updates the weights of an MLP in-context under an auxiliary key→value objective, but the writer is still largely fixed: the projections for keys/values and how strongly to update (how fast to learn, how much to forget) are baked in by pre-training. Hope loosens that constraint. It replaces those fixed projection-and-update rules with additional learned memory modules that start from pre-trained weights and then update in-context, producing the parameters of the write step—keys, values, and per-step learning/forget controls—on the fly, and updating itself with a Delta rule based optimizer. Thus, Hope swaps self-attention’s non-parametric “read from the entire context” mechanism for a parametric associative memory system whose encoding and update behavior can adapt as the context evolves.

Continuum Memory System (Replacing Static MLPs): Standard Transformer MLPs are frozen after pre-training. Hope replaces them with a chain of MLP blocks updated at different frequencies—higher-frequency blocks update more often (on shorter token “chunks”), while lower-frequency blocks update more slowly. This is motivated as a memory-consolidation analogue: new information is captured in fast, plastic components and can become reflected in slower, more stable components through the multi-timescale learning dynamics/knowledge transfer. The idea is that blocks influence each other through their coupled computations and shared training dynamics, so information represented in faster-changing components can, over time, be reflected in slower ones and help reduce catastrophic forgetting. Mechanistically, each block holds its weights fixed within a chunk and then applies a single update at the chunk boundary using the accumulated optimizer update signal over that chunk.

Experimental Results

The results suggest that bridging the gap between context and weights unlocks meaningful continual learning capabilities.

The experiments most directly applicable to current systems evaluate Hope-Attention: the CMS multi-timescale memory system combined with standard Transformer attention, rather than the fully attention-free Hope architecture. In these experiments, they take a pretrained Llama 3 backbone and replace the frozen MLPs with CMS, then continue training for 15B tokens.

On class-incremental learning benchmarks, where models must sequentially learn new categories without catastrophic forgetting, this Hope-Attention variant outperforms both standard in-context learning and a number of other continual learning methods. The interpretation is that multi-timescale parametric memory enables knowledge retention that prompting or attention cannot replicate.

Long-context QA/retrieval shows consistent benefits from adding memory levels: adding levels and updating the most persistent CMS level more frequently improve performance on needle-in-a-haystack and QA benchmarks, though with a compute tradeoff.

Perhaps most striking is the 'Continual Translation of a Novel Language' experiment, where models must learn two low-resource languages—Manchu and Kalamang—sequentially in-context. While standard in-context learning suffers catastrophic forgetting of the first language once the second is introduced, adding CMS levels progressively restores retention, with the three-level variant nearly recovering the original single-language baseline. This suggests that the adaptive CMS blocks effectively offload in-context adaptation duties from the attention mechanism, allowing the model to compress and persist linguistic rules within its weights rather than relying solely on traditional attention based in-context learning.

Training Hope from scratch (as a fully attention-free backbone), they show Hope outperforming baselines, including Transformers and Titans, on standard language modeling and common-sense reasoning benchmarks while displaying better scaling properties than other attention-free models. On short-context recall tasks, and Needle-In-A-Haystack experiments, Transformers with standard attention still outperform, but Hope improves on other attention-free options.

On BABILong, a fine-tuned version of Hope significantly improves on Titans and reports greater than 90% accuracy at 10M tokens sequence length, while large models like GPT-4 fail at around 128K–256K tokens. They note that the fine-tuning is necessary for the slowest updating levels to adapt, allowing the faster updating layers to compress 10M tokens “in-context”.

Like Titans, Hope achieves O(1) memory in context length: the main long-range state lives in fixed-size weights rather than an ever-growing KV cache. The tradeoff is a constant-factor compute penalty for the write and consolidation steps that Transformers skip at inference. As with Titans, this involves chunk-wise gradient computation and optimizer state for the updating components.

The infrastructure challenges for per-user deployment are more severe than Titans. Where Titans updates only a memory MLP, almost all of Hope's parameters can be updated per-session. Hope is probably best understood as a proof of concept for these building blocks, demonstrating that weight-based updates can augment, replicate or improve on the short-term non-parametric memory of attention. Both CMS blocks and self-referential Titans modules are compatible with fixed-window attention, and it is unclear how exactly they might best be integrated into future frontier models.

The largest experiment reported with the Hope architecture is with a 1.3 billion parameter model trained on 100 billion tokens. For the Hope-Attention experiments they use the Llama-8B and continue training it with the MLP blocks replaced with their CMS block for 15B tokens.

Near-Term Applications

It remains unclear whether the techniques described here, or other weight-based continual learning approaches, will scale to solve the problem properly. But I find them promising, and I strongly suspect several AGI labs are actively working on finding the optimal configuration, training recipe, and integration with existing RL pipelines enabling weight based continual learning at scale.

The specific continual learning method that ultimately succeeds may end up looking quite different from Titans or Hope. But any solution achieving genuine weight-based learning will probably face similar constraints: inference-time gradient computation and per-user weight divergence breaking the efficiency of batched serving.

For standard chat interfaces with limited context needs, traditional LLMs will likely remain dominant. The infrastructure overhead of per-user weight updates during inference makes broad consumer deployment challenging in the near term. Initial applications will more likely emerge where weight updates can be computed offline and shared across many users, amortizing the overhead across many requests served with standard inference. Weight-based memory could also prove valuable wherever cached KV-prefill is already worthwhile, with the advantage that learned weights are fixed-size regardless of how much context was compressed into them.

For model providers, if these techniques avoid catastrophic forgetting without introducing unpredictable behavioral changes, they could enable continual training on recent news and world events, keeping models current without full retraining cycles. The result would be models that simply know the latest news rather than needing to search for it.

The most interesting possibility may be enterprise applications. At sufficient scale, it makes sense to persist weight updates rather than recompute them each session. I can imagine services where companies pay to maintain continually updated custom weight checkpoints trained on their codebases, internal documentation, public Slack messages, and relevant industry news, producing an assistant with more institutional context than any individual employee. Most work with chatbots today is just providing them with the context they need; weight-based memory could make that context persistent by default.

The storage and infrastructure overhead for maintaining custom model checkpoints wouldn't be trivial. Depending on architecture, this might only require updating a subset of weights. Titans updates an extra memory MLP, while other approaches have explored training some subset of Transformer weights^[2]. But you’re still essentially versioning and serving personalized model weights rather than just prompts and context or a tiny LoRa.

However, if the result is something approaching a superhuman remote knowledge worker, one that genuinely internalizes your codebase, processes, strategic context, and tacit organizational knowledge, faster than a very smart and experienced new employee would, the economics could still work out decisively in favor. Companies pay $200k+ for a single senior engineer. Even a single multi-day agentic task might justify maintaining unique model weights for the duration of the trajectory: current context compression is very lossy, while weight-based memory may offer far greater capacity and learned selectivity about what to retain.

Timelines implications

What would solving continual learning mean for AI progress? METR's work on time horizons provides a useful lens. They define an AI model's 50% time horizon as the length of tasks (measured by human completion time) that it can complete autonomously with 50% probability. On a diverse benchmark of software and reasoning tasks, this metric has doubled roughly every 7 months over the past 6 years—with evidence suggesting an acceleration to approximately 4-5 months^[3]from 2024. Many attribute this acceleration to RL scaling. The RL trained reasoning models that emerged in late 2024 demonstrated that RL post-training could unlock substantial capability gains beyond what pretraining alone achieved.

Weight based continual learning could represent a similar inflection. Just as RL let models "think longer," continual learning could let them think longer and remember more in a way that is not limited to the size of the context window. Current scaffolding approaches hit friction in long-horizon tasks: repeated context compression^[4], retrieval failures, accumulated errors across summarization cycles. A model that genuinely learns from experience could sidestep these bottlenecks, maintaining coherent state over days or weeks without the very lossy context compaction. Models could also learn from deployment itself, accumulating task-specific expertise that compounds over time.

Safety implications

Many current alignment techniques assume frozen weights. RLHF and constitutional AI shape behavior during training; red-teaming probes for failures before deployment. But once shipped, the model is static. Weight-based continual learning breaks this assumption. Safety guarantees are no longer fixed at deployment; they can drift through deployment interactions in ways that are difficult to monitor or predict.

At the technical level, guaranteeing safety properties for every input to a fixed function is already hard; guaranteeing them for a function that modifies its own weights during deployment is qualitatively harder. The attack surface evolves with use, and the verification problem has no clear stopping point.

At the conceptual level, the concern is ontology shift. Alignment techniques like constitutional AI train behaviors relative to the model's current internal representations. If continual learning shifts how the model categorizes concepts (what counts as "deception," who counts as a "person," which actions constitute "harm"), the same trained behaviors may produce different outputs. A refusal that fires reliably for a frozen model may not fire once the model has learned to frame the same request differently.

I recommend Seth Herd's post "LLM AGI will have memory, and memory changes alignment" where he develops some of these concerns in more detail.

If continual learning does become standard, this likely shifts emphasis from pre-deployment evaluation toward runtime monitoring, constraints on what the online learning process can modify, and incident analysis when failures occur. It probably also strengthens the case for policy as a safety lever.

If continual learning is genuinely a key bottleneck, solving it could produce a step change in general capability as models suddenly accumulate skills across long-horizon tasks, learn from deployment, and compound their own improvements. We'd be introducing a new learning paradigm with poorly understood alignment properties at precisely the moment models become substantially more powerful.

Conclusion

A lot of alignment work assumes frozen weights, and it's unclear how much of it transfers if that assumption breaks. If you work on alignment, this area deserves more attention than it's currently getting. And if you're at a frontier lab and it starts looking like continual learning is on the verge of being solved, some advance notice would be appreciated.

1. Humans with anterograde amnesia retain procedural learning. They can improve at motor skills like mirror tracing or puzzle assembly across sessions without any conscious memory of having practiced. ↩︎

2. Another recent weight-based online learning technique, E2E-TTT, modifies only the final quarter of its transformer MLPs between chunks, needing only to compute chunk-wise gradients for the last quarter of the network. ↩︎

3. See METR’s 4 month claim, Peter Wildeford finding 4.4 months on 50% reliability , and Ryan Greenblatt predicting 5 months for the next 2 years. ↩︎

4. Opus 4.5, currently the model with the longest 50% task time horizon, has a maximum context window of 200k tokens, and METR uses a token limit per task of 8 million tokens. ↩︎