A model can only memorize so many facts in its weights — so how do you get past that ceiling?
Can in-weight memorization scale beyond model parameter count limits?
This explores whether a model can hold more facts in its own weights than its size allows, and if it can't, how the corpus says to get around that ceiling.
This explores whether a model can hold more facts in its own weights than its size allows, and if it can't, how to get around that ceiling. The corpus says no: in-weight memorization is capped by model size. One formal proof plus experiments shows that in-weight recall is bounded by parameter count, while a simple tool-use circuit lets factual recall grow without that bound Can models store unlimited facts without growing larger?. The cap can be measured. GPT-family models store roughly 3.6 bits per parameter, and this is a property of the model, not of the training method When do language models stop memorizing and start generalizing?.
The ceiling is more than a limit, because it is also where the model's behavior changes. When memorization capacity fills, models shift from memorizing to generalizing, a transition known as grokking When do language models stop memorizing and start generalizing?. So parameters are a shared budget, and stored facts compete with general skill for the same room. Forcing more facts in with finetuning is worse than hitting a wall, because it degrades general capability by overwriting prior knowledge Can models store unlimited facts without growing larger?. Memorization also takes up physical space. Memorized paragraphs leave a fingerprint in the low layers, including a specific attention head that attends to rare tokens Where does a model store memorized paragraphs?.
Most of the escape routes leave the weights. Tool use is the cleanest, since the facts live in a database and the model only needs the lookup skill. A separate memory model is a middle path. It is trained to encode new knowledge, works with frozen proprietary LLMs, and keeps inference cost from growing with corpus size, though it has capacity limits of its own Can a separate memory model inject knowledge without touching the LLM?. Titans-style neural memory modules change what memory is. A long-term memory prioritizes surprising tokens and sits alongside attention, which lets the model handle 2M+ token contexts without quadratic cost Can neural memory modules scale language models beyond attention limits?. That stretches how much the model can keep in view, not how many facts its base weights hold.
The one route that works on the weights themselves is the sleep paradigm, which consolidates in-context knowledge into weights through Knowledge Seeding, a step that distills a smaller network into a larger one, plus RL-generated rehearsal Can models consolidate memories during offline sleep phases?. It doesn't break the ceiling. Moving knowledge into a bigger network is what the bound predicts you'd need. Looped models are often mentioned alongside these. They gain reasoning through repeated depth instead of extra parameters Can models learn by looping instead of growing larger?, but that addresses a different limit. It concerns reasoning compute, and the corpus doesn't claim looping raises fact-storage capacity.
Moving memory outside the weights has its own cost. In MemTrapBench, all five memory frameworks underperformed a no-memory baseline by more than 10%, even though the stored memories were accurate and relevant Can relevant memories actually harm LLM reasoning?. Leaving the weights removes the storage ceiling, but the hard problem becomes using what's stored without being misled by it.
Sources 8 notes
A formal proof and experiments show in-weight memorization is bounded by model size, while tool-use enables unbounded factual recall through a simple circuit. In-weight finetuning also degrades general capability by overwriting prior knowledge.
GPT-family models have a measurable memorization capacity of approximately 3.6 bits-per-parameter. When this capacity fills, a phase transition triggers grokking—the shift from memorization to genuine generalization. This capacity is a property of individual models, not training algorithms.
Memorized paragraphs leave a distinctive fingerprint in GPT-Neo: larger gradients in lower layers, concentration in a specific low-layer attention head attending to rare tokens, and dependence on a few early-prefix tokens. This localization makes memorization targetable for unlearning.
MeMo trains a dedicated memory model to encode new knowledge, eliminating inference-time search costs that scale with corpus size. It avoids fine-tuning risks and works with frozen proprietary models, but trades this for up-front training cost and capacity limits.
Titans architecture separates attention (short-term, quadratic) from neural memory (long-term, compressed), prioritizing surprising tokens for storage. The model outperforms standard Transformers and linear RNNs across tasks while scaling to 2M+ token contexts without quadratic penalties.
Show all 8 sources
The Sleep paradigm uses Knowledge Seeding (distilling smaller networks into larger ones) and Dreaming (RL-generated rehearsal) to consolidate in-context knowledge into weights without forgetting. Gains appear in long-context understanding, few-shot reasoning, and continual learning.
Models that re-apply layers in recurrent depth outperform larger feedforward networks on reasoning tasks. This works because recursion enables state tracking and compositional generalization that parameter scaling alone cannot achieve, with convergence signals providing natural halting.
MemTrapBench shows that all five tested memory frameworks underperform a no-memory baseline, with drops exceeding 10%, despite memories being accurately stored and task-relevant. This reveals a failure mode at the point of use that standard memory benchmarks miss.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Diagnosing Memorization in Chain-of-Thought Reasoning, One Token at a Time
- Continual Learning Mechanisms Compose for Long-Horizon Memorization
- Provable Benefits of In-Tool Learning for Large Language Models
- MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use
- How much do language models memorize?
- Memorization and Knowledge Injection in Gated LLMs
- Localizing Paragraph Memorization in Language Models
- Know It, Act on It: Investigating Memory Utilization in LLM Personalization