I haven’t drafted more than a sentence of this blogpost, but I’m already preparing to bark at ChatGPT that I never use colons when writing…for the 50th time this week. I’ll probably also have to tell it to simplify its description of key mathematical concepts (I have, afterall, been in VC for 10 years) at least 10 times when I review ICML papers later this week. All of this is to say, my interest in memory and continual learning really stemmed from personal experience. My local barista never remembers my coffee order (decaf Americano with a splash of oat milk), but I hoped my AI would.
It turns out I’m not the only one. I soon turned to my friends and colleagues, including researchers and engineers building GenAI products and applications, to find out whether they shared similar interests.
- Did the people building coding agents want AI that would learn when to add more explanatory comments and when to keep them minimal?
- Would the people building legal agents want AI that would learn when to write brief summaries and when to produce a comprehensive, citation-rich analysis?
- Would customer support teams want AI that could adapt to a company's preferred tone, level of formality, and escalation practices?
Should AI learn not just what we want it to do, but how we prefer it to do it?
The answer was a resounding yes. Because AI engineers do, in fact, want to build products that you’ll like and then love. But also, because they don’t want you to switch to a competitor. You’re more likely to continue using a product that gets better as you interact with it. At a time when distribution is cheap, retention really matters.
They had other concerns, too. Many were relentlessly focused on releasing agent skills. While their agents could compose these skills to automate a handful of common but tedious workflows, they struggled to build and maintain enough skills to address the long tail of workflows their customers ultimately wanted automated. They really wanted to build agents that could learn new skills based on past experiences.
At the center of all these conversations were two themes: memory and continual learning.
My next step was to discover what, if any, techniques might solve this problem, so I began speaking with researchers in industry and academia. At first, many people pointed to context engineering. Just load conversation logs into the context window! Cram all the search results into the context window! Oh, the context window is too small? Don’t worry, we’ll just figure out how to train longer context windows.
But it turns out that long context pre-training is really hard. Potentially so hard that it might not be possible to do with Transformers. And even if you could do it, it would likely require much bigger models and be prohibitively expensive. What’s more, cramming lots of data into context is a great way to spend a lot of money only to produce bad answers. Have you heard of context rot? Even if you haven’t, you’ve probably experienced it. Some studies have empirically demonstrated that as the number of examples you put into context grows, the model accuracy declines.
To repeat the refrain of many a blog post…there had to be a better way! And that’s when I read about Cartridges. Sabri Eyuboglu and a team from Stanford published their research on CartridgesCartridges: Lightweight and general-purpose long context representations via self-studyLarge language models are often used to answer queries grounded in large text corpora (e.g. codebases, legal documents, or chat histories) by placing the entire corpus in the context window and leveraging in-context learning (ICL). Although current models support contexts of 100K-1M tokens, this setup is costly to serve because the memory consumption of the KV cache scales with input length. We explore an alternative: training a smaller KV cache offline on each corpus. At inference time, we load this trained KV cache, which we call a Cartridge, and decode a response. Critically, the cost of training a Cartridge can be amortized across all the queries referencing the same corpus. However, we find that the naive approach of training the Cartridge with next-token prediction on the corpus is not competitive with ICL. Instead, we propose self-study, a training recipe in which we generate synthetic conversations about the corpus and train the Cartridge with a context-distillation objective. We find that Cartridges trained with self-study replicate the functionality of ICL, while being significantly cheaper to serve. On challenging long-context benchmarks, Cartridges trained with self-study match ICL performance while using 38.6x less memory and enabling 26.4x higher throughput. Self-study also extends the model's effective context length (e.g. from 128k to 484k tokens on MTOB) and surprisingly, leads to Cartridges that can be composed at inference time without retraining.arXiv:2506.06266v3View paper in June 2025. They described an approach in which a small KV cache could be trained offline through self-study on a corpus of documents (e.g., conversation logs, code repositories, or sets of long documents). At inference, this Cartridge could be loaded to produce better outputs.
You’d no longer need to stuff the same documents, codebases, and knowledge into the prompt every time; you could pre-compute it once and save it as a compact memory object that the model would load when needed. It seemed clear that this research would be very meaningful (and fortunately, I knew the author - but more on that later) - it was cheaper, faster, and more scalable than traditional RAG and would mean that the model, unlike my barista, never forgot about past interactions with me. I knew this could become a fundamental building block for persistent memory and efficient context in AI systems.
Although Cartridges are promising for solving personalization and many other practical problems in AI, I still believed we would need other strategies to update model weights so it could draw connections between its new insights and knowledge acquired during pre-training. It seemed crazy to pre-train models once and then adjust their weights only through RL post-training, a setup designed to augment existing capabilities rather than cultivate new ones. The human brain changes constantly as it accumulates new memories through synaptic remodeling and neural pathway evolution; surely the agent brain couldn’t just be static.
I started looking into parametric memory updates and quickly found research by Jessy Lin and a team from Meta and Berkeley on sparse memory fine-tuningContinual Learning via Sparse Memory FinetuningModern language models are powerful, but typically static after deployment. A major obstacle to building models that continually learn over time is catastrophic forgetting, where updating on new data erases previously acquired capabilities. Motivated by the intuition that mitigating forgetting is challenging because trainable parameters are shared across all tasks, we investigate whether sparse parameter updates can enable learning without catastrophic forgetting. We introduce sparse memory finetuning, leveraging memory layer models (Berges et al., 2024), which are sparsely updated by design. By updating only the memory slots that are highly activated by a new piece of knowledge relative to usage on pretraining data, we reduce interference between new knowledge and the model's existing capabilities. We evaluate learning and forgetting compared to full finetuning and parameter-efficient finetuning with LoRA on two question answering tasks. We find that sparse memory finetuning learns new knowledge while exhibiting substantially less forgetting: while NaturalQuestions F1 drops by 89% after full finetuning on new facts and 71% with LoRA, sparse memory finetuning yields only an 11% drop with the same level of new knowledge acquisition. Our results suggest sparsity in memory layers offers a promising path toward continual learning in large language models.arXiv:2510.15103v1View paper. They described an approach, enabled by parameter-efficient fine-tuning with LoRA, that updates only memory slots activated by a new piece of knowledge, thereby minimizing catastrophic forgetting. They showed empirically that this approach could support new knowledge acquisition with much less forgetting. This was the missing component.
At this point, I knew what I was looking for: a platform that could enable personalization, minimize the need for search and retrieval, and support the development of new skills - one that was designed around KV cache compaction to leverage new knowledge at test-time but also sparse memory fine-tuning to support continual learning.
Fortunately, the opportunity landed in my lap. You see, over the course of 5 years, I have met 3 remarkable individuals:
I already mentioned Sabri, whom I first met in 2021 through his work on Meerkat, a library that makes data inspection, evaluation, and training of multimodal and other models more efficient and robust. Sabri is one of those rare researchers who combines extensive knowledge of coding and systems with a nose for novel but practical research ideas. He also taught me the word “sus.”
At an Amplify dinner in 2022, I met Jessy Lin (also previously mentioned), whose obsession with and knowledge of research on continual learning and memory still exceeds anyone I’ve met after 10 years of connecting with researchers.
In 2023, I also met Jack Morris in his second year as a PhD student at Cornell. While he told me that he didn’t plan to continue his research on deidentification in ML, he later shared his conviction that better approaches to memory would enable the development of better consumer and enterprise products. And his thesis on memory in language models was a banger.
But it wasn’t until 2025 that I met Dan Biderman, who corralled this incredible group of memory researchers together to commercialize their work and change the way AI agents are designed. I was familiar with Dan’s researchLoRA Learns Less and Forgets LessLow-Rank Adaptation (LoRA) is a widely-used parameter-efficient finetuning method for large language models. LoRA saves memory by training only low rank perturbations to selected weight matrices. In this work, we compare the performance of LoRA and full finetuning on two target domains, programming and mathematics. We consider both the instruction finetuning (approximately 100K prompt-response pairs) and continued pretraining (20B unstructured tokens) data regimes. Our results show that, in the standard low-rank settings, LoRA substantially underperforms full finetuning. Nevertheless, LoRA better maintains the base model's performance on tasks outside the target domain. We show that LoRA mitigates forgetting more than common regularization techniques such as weight decay and dropout; it also helps maintain more diverse generations. Finally, we show that full finetuning learns perturbations with a rank that is 10-100X greater than typical LoRA configurations, possibly explaining some of the reported gaps. We conclude by proposing best practices for finetuning with LoRA.arXiv:2405.09673v2View paper defining best practices for fine-tuning with LoRA. But I didn’t know how special he was until something remarkable happened. Jonathan Frankle, who has never said anything nice about anyone, said he was good.
I would have backed anything this team developed, but it just so happened they decided to focus on the one thing I believed builders needed most. A way to r̶e̶m̶e̶m̶b̶e̶r̶ ̶m̶y̶ ̶c̶o̶f̶f̶e̶e̶ ̶o̶r̶d̶e̶r̶ make agents useful and delightful through better memory management and continual learning. The Engram team has the right people (Dan, Jack, Jessy, Sabri, Scott), the right approach (initially focused on Cartridges but expanding into sparse memory fine-tuning), and an unrelenting focus on making customers like Notion, Harvey, and Microsoft successful.
That’s why today I am excited to share that Amplify is part of Engram’s $98M in funding, announced today via CNBC.
My AI apps still don’t remember that I ONLY EVER WANT TO SEE MEDICAL RECOMMENDATIONS BACKED BY RCTS or that I WILL NOT STAY AT HOTELS WITH CARPET. But Engram will change that soon. So you should probably get in touch.
