Memory and Continual Learning: Engram's Dan Biderman and Jessy Lin
In a Nutshell
Engram's core claim is that the next leap in usefulness comes from internalizing private context directly into model weights via continual adapter fine-tuning, rather than relying on external memory or ever-larger generic models. By compressing company documents, workflows, and interactions into compact LoRAs or similar adapters, inference cost drops up to 100× while the model gains the implicit knowledge of a long-tenured employee. The approach treats training as an always-on process that decides what to bake in versus leave in retrieval, solving the scaling limits of token-heavy context engineering.
These notes were generated by AI and may contain inaccuracies.
The fundamental question is what about pre-training or post-training enables models to generalize in emergent ways, and how to control that process so a company can make models learn private data just as deeply as they know the capital of France or how to write Python.
Engram is a new lab focused on memory and continual learning. The website states they do not see the world through pre-training or post-training. Their models are always training. The bottleneck for making models more useful is not raw intelligence but understanding new and evolving context, whether a new task or a particular job context. The goal is to bake that context into the model weights the same way pre-training and post-training do.
Memory as a separate database or context window is not true memory or true continual learning. Context engineering, tool use, and large prompts help but are insufficient. The underleveraged approach is applying the same training pipelines frontier labs use for math or code to every domain and private company context. As individuals, taking notes and using sticky notes is valuable, but people also carry internalized traces of memory and new intuitions into the next day. Current solutions are more externalized memory.
The amount of tokens generated collectively will reach tens of millions per day soon. Keeping, searching, and rereading this volume will be expensive and confusing for models without major breakthroughs. The vision is combining externalized memory with internalized traces.
Engram works with partners like Notion, Microsoft, and Harvey. These workspaces contain documents written by teams and conversations with agents over long periods. The goal is a model that deeply understands context the way a long-tenured employee does, grasping company initiatives, ways of working, hiring pipelines, and priorities. Rather than reading files only at test time, the model internalizes this context.
Training happens into the weights using adapter fine-tuning of many types including LoRAs, prefixes, and sparse architectures. The second challenge is determining the right data and turning raw documents or interactions into useful training signals. Tools include supervised fine-tuning, RL, and on-policy distillation. The premise is not that tools are absent. Some knowledge remains externalized and tools are always available. The hard task is deciding what needs to be internalized versus externalized. Companies have bespoke tools and workflows that frontier models were not trained on.
A Notion agent becomes a custom agent that is LoRA fine-tuned or adapter-tuned to continuously learn on new content added to a workspace. This works on open-source models easily. It can work with closed frontier models if white-box access to weights is available. Any transformer model can be adapted.
The trade-off involves burning more compute upfront to learn company ways of doing things into the weights, then sending less context on every inference pass. Benefits include not needing to research or reread things and not writing monstrous system prompts. This can yield two orders of magnitude reduction in token inference consumption, potentially 100x fewer tokens. Things relating to people, teams, organization, and priorities are hard to find in single documents. The model can implicitly learn these and answer in 100 tokens what frontier models would consume 100,000 tokens to handle. There is a consistent three-to-six-month gap where bespoke tasks are not yet handled well by general models, and autonomous lightweight learning provides value during that window.
Either workspace or individual level works. It is easier to start with teams because they are more disciplined in collecting context and have larger accumulated context over years. Every person's computer and phone will eventually be a target. The largest deposits of information are currently in teams collaborating on knowledge work.
Large language models have substantial fact memorization built in. Some argue models should abstract concepts like countries and capitals rather than memorize specific facts. The question is how to disentangle memorization versus learning. Remembering facts is necessary to compose them into more complex concepts. What is missing is figuring out what is important to remember. Academic benchmarks often test learning very specific facts, such as the length of a bridge in an African country, that models should not devote capacity to. Human memory is lossy because intelligence involves compressing what is important and separating it from what is not. Fact learning cannot be cleanly separated from skill learning. Stripping facts from a model leaves it unable to perform basic operations needed for further reasoning.
Efforts to disentangle algorithms from databases and pre-train models to retrieve rather than internalize facts have been hard to scale. The current recipe that works is fact pre-training. Deep learning has mushed together what traditional computer science kept separate: databases storing facts and algorithms manipulating information. Interpretability research tries to separate them again. In economic adoption of AI, companies handle their own contexts carefully while generic models operate on them as strangers. There needs to be convergence where facts, stories, and details get mixed into the model. A disadvantage is that facts change, and imprinting them into weights is challenging.
The key distinction is not algorithms versus databases but how to distinguish what is important to remember versus what is not. Inspiration from human dreaming is taken loosely. Dreams involve a phase of deeply internalizing context that is currently missing. Everything now happens at test time with thinking on the fly, which leads to mistakes. The goal is to digest experience back into the model so future performance improves.
Dreams are incoherent but involve experimenting with affordances in the world, social situations, and heavily biased toward social content. The analogous process for models is giving them time to retreat from interaction and experiment with what they can do in an environment, what they know, and how to handle tail extreme cases.
A canonical motivating example is an AI lab needing to win a math Olympiad in one week. The superior approach is not constructing annotated catalogs of textbooks but collecting and synthesizing training data, launching a training job, evaluating, and iterating. This magic of training needs to extend beyond high-stakes domains like math, coding, and cyber into many more hands.
The worldview differs from frontier labs, which aim for one ever-larger model intelligent across domains. Instead, the vision is everybody having their own model. Many things people want to learn are private or never appear in post-training datasets, or are conflicting across companies and individuals. Many real-world tasks are ambiguous with no clear ground-truth reward signal, unlike clean supervised environments used in frontier training. These tasks are specific to individuals and misaligned with how frontier labs structure their pipelines.
Frontier lab P0 is reaching AGI with one generic model extremely capable in coding and math, then using it to automate the economy or solve hard problems. Memory and continual learning are on their radar but treated more as product efforts. Dedicated focus and breakthroughs are needed. At a Sequoia event, Demis Hassabis stated new breakthroughs around these topics are required. Incentives around where data lives and who owns the model matter. Learning from many humans or organizations at scale without shoulder-to-shoulder work would be a significant unlock.
Three changes are needed: new research breakthroughs, new infrastructure for training small models for everybody rather than single large runs, and a different integration of research and product. In a world where models are always training, user inputs are intricately tied to training signals. An integrated loop between research and product is required, with rapid deployment to gather real-world feedback.
With models already very smart, the remaining challenge is learning at the edges, the remainders that make models useful. This is not just raw intelligence but learning new things. It feels fundamental because it requires understanding why pre-training works so well. The internet provided a diverse gift of data containing examples of coding, writing, and other domains. The task is figuring out what about pre-training enables magical generalization and controlling that process for private company data so models learn it as deeply as they know capitals or Python.
Dan Biderman came from neuroscience, initially interested in consciousness and the human condition, then in how humans think and perceive. He was drawn to statistics and AI where memory and continual learning problems are urgent. Current AI solutions are far from biology. Human memory evolved under information capacity constraints to have fuzzy representations that support abstraction, connection formation, and next-day use. Current systems lack this beyond generic pre-training. The interest is in building analogous capabilities without making machines exactly like brains, since computers excel at lossless storage of entire codebases.
Modern computer architecture has CPUs, GPUs, and memory as different co-processors. The question is whether LLMs will converge to one dominant co-processor or build multiple co-processors emergently, and whether memory will be an emergent property or require standalone architecture.
On the model architecture level, there have been many previous experiments with different architectures, including the state space family, to try and handle very long context more efficiently. All these methods end up being a trade-off between memory and accuracy. The transformer attention mechanism has quadratic memory footprint in sequence length, while these alternative models claim sub-quadratic costs. Some of the best Chinese models have layers inspired by state space architectures and are not quadratic in cost. However, in practice, there is always a compromise of accuracy for memory savings.
The approach is to burn more compute on new context that has not been seen before rather than optimizing for memory efficiency at the cost of accuracy. The direction of AGI is not expected to end anywhere soon, and there is more compute to scale. Understanding specific contexts requires more than just rereading files.
People are spending time in the app layer trying to make models work well for specific use cases, such as designing a website with a particular brand style, learning how to run a workflow, or adopting a particular way of writing. There is a distinction between facts and skills, though the methods are agnostic to that distinction.
At the app layer, frontier models are wrapped in loops with search tools. The interest is in interfacing with data in ways that are faster, more efficient, and more contextual. This includes encoding firm knowledge in something efficient that does not require repeated research, and having the model know in a targeted way who to triage things to. With lightweight training, these things can become instinctual to models rather than requiring involved long inference loops.
This functions as a RAG killer in the sense that while retrieval can always be done, people are interested in interfacing with very large data planes and automating repetitive things through model weights.
The question of what should be done in model weights versus externalized to RAG is an unsolved problem. It is also the fundamental question of biological memory—what should be internalized versus what should not. A hotel room number from a year ago does not need to be internalized in neural tissue, but a current home password might be useful to have imprinted for the next few years.
The approach is to use as few heuristics as possible. It is easy to run filters on data and decide what to keep or discard, but humans are exposed to a lot of garbage and still learn without going off the rails, and models should be able to do the same.
In the short term, a major pain point is the high inference costs of running agents for days on end. The same files are being read over and over again, even across people in the same company running the same queries on the same documents. This should be something the model just knows, similar to how an employee knows what they were working on yesterday without typing into a search box.
Caching solves this to some extent, but the question of what should be internalized versus retrieved at test time remains. If retrieval is always used, associations cannot be made at an abstract level. For example, seeing someone on a team doing certain research and recalling a related thing that might be useful, without being explicitly asked. These kinds of associations can only happen in weights.
The main limitation with retrieval systems is not what to store or where to put it, but how to address it—how to query the thing. Models sometimes lack the intuition to know where to look, especially when limited to keyword search, which is easier to scale in terms of infrastructure for embeddings.
Knowing what to search for is something intuitive that can happen in the weights. Much of the work started with a deep dive into KV caches. A KV cache for a single Wikipedia article can be around 80 gigabytes of HBM memory on the GPU. For a 70B Llama model, the entire weights would be about 100 gigabytes. The KV cache takes tens of kilobytes of an article and expands it to 80 gigabytes of brain state.
Gradient descent can pack a lot of information in very few numbers, whereas KV cache is inefficient. The idea is to take those 80 gigabytes, spend compute offline to compress it, and make the thing loaded in cache a thousand times smaller. This would have tremendous implications for how things are loaded, how fast operations can be performed, and the fidelity of the representation.
A proof of concept people are waiting to see with continual learning is having an intern that can be taught things over time and actually gets better. No matter how sophisticated context engineering approaches are, they are not getting there. The expectation is that the model is actually getting smarter in a way that is different from yesterday.
The original ChatGPT model was not anticipated, and many product directions existed before it. If someone resigned from their job today with the sole mission of making a model better for themselves using frontier models and engineering context 24/7, the ability to move the needle would be very limited. It would be better to wait for the next version of the model. The goal is a future where time spent on a model translates to quality of performance in domains that matter.
This is hard to achieve and requires scaling compute and training on data without destroying it.
Recent surprises include the capabilities of coding agents, though there had been prior effortful experimentation with these systems. The main events identified are GitHub Copilot, ChatGPT, and seeing agentic capabilities. There is anticipation around how far agentic systems can go and what long horizon they can achieve.
Models changing in a way that is not harmful and learning new things on the fly in a personally and economically viable manner is the interesting development.
There is discussion around whether there will be a memory bank or memory wallet that moves across the digital world. The right product form factor is not fully figured out. With ChatGPT memory, there is reluctance to have it remember across personal and work contexts, as cross-contamination would be irrelevant and undesirable. Memory in tools and products should be separated to maintain control.
A holy grail would be burning through tokens at work to create value while IP stays with the company, but skills learned, things invented, and ways of doing things can be taken to the next job in a sanitized way that is not harmful to other companies' IP. Carrying a set of skills would be interesting. This happens in biology with NDAs and ethical rules, and doing it in the digital world would be rewarding because it would force deeper implementation of AI and reward that effort.
Starting a PhD in 2007, AI was focused on statistical learning with computer vision and NLP as the two main areas. In 2012 with AlexNet, vision dominated for about six years. It is surprising that language has come to dominate over vision in progress.
Language is interesting as a medium for communication where complex abstract things can be done. In the longer term, language and vision will likely combine in a unified system taking inputs from multiple modalities and understanding them abstractly.
There was a bias toward vision as the natural direction, with the expectation that experiencing the world through action and vision would be key. After the ChatGPT moment and work on the NLP side, it became striking how well language works despite each word having a one-hot embedding vector in a high-dimensional space that is completely dissimilar to other words. Models that are an order of magnitude bigger than the best vision models still make things work well.
There is still a lot of juice to be squeezed in image and video domains.
A theory proposed is that in biology, vision has a massive fundamental advantage over language because the bit rate the brain can process through the eye is many orders of magnitude greater, with optical processing happening before electrons reach the brain. The total bit rate of training data processed and reaching the brain is many orders of magnitude greater than audio data, where sound waves are fundamentally slower.
There is an upscaling from acoustics to electronics in hearing, while there is downscaling from photons to electrons in vision. In computers, everything is electronic, so vision is nerfed and language is promoted because all processing is on the same playing field. This might explain from an information theory perspective why language and vision end up on a similar playing field with LLMs, and why the architecture is better suited for language than vision.
In the limit, everything is RAG. Only important things are internalized. When training vision-language models, the language component ends up dominating the vision content.
A certain brain being more biased toward a modality does not necessarily mean processing will be more efficient for that modality. Efforts on brain-computer interfaces should account for what real estate exists in the brain for relaying information back. For knowledge work, text is equally fine.
If everything goes right, the world in five to ten years would have everyone with their own model that is really different from other people's models and from frontier models. These serve different purposes, with models that know you and help in ways that make sense personally, whether for an individual or a team. There would be different kinds of intelligence everywhere.
In neuroscience, memory and navigation are closely related, with the same circuits representing landmarks in space also handling elements of episodic memory. The company could serve as the actual LLM interface to the data plane for everyone, forming neural memories that are personalized. There would be hundreds of millions of such models representing a brain state of the file system rather than the file system as it is. This would be more efficient and associative than current approaches.
Keep Sequoia Capital in your library
Save the videos and channels worth coming back to, and find them again in one place.
