# Engram (company)

> Source: https://aiwiki.ai/wiki/engram_ai
> Updated: 2026-10-27
> Fact-checked: 2026-10-27
> Categories: AI Companies, AI Research, Enterprise AI, Large Language Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Engram (company)." aiwiki.ai, 27 Oct 2026. https://aiwiki.ai/wiki/engram_ai
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Engram** is an American artificial intelligence company, based in San Francisco, that trains [large language models](https://aiwiki.ai/wiki/large_language_model) to absorb an organization's documents, messages and other working context into their weights. The company calls the result a "learned memory layer." Instead of re-reading the same files on every query, an Engram model studies a workspace ahead of time and answers from what it has learned, which the company says lets it match or beat frontier models while using up to 100 times fewer tokens.[1][4] Engram was founded in October 2025 by Dan Biderman (chief executive), Sabri Eyuboglu (chief technology officer), Jessy Lin, Jack Morris, Scott Linderman and [Christopher Ré](https://aiwiki.ai/wiki/christopher_re), researchers with ties to Stanford, Berkeley and Cornell.[1][2][3] It emerged from stealth on June 23, 2026 with $98 million in funding from investors including [General Catalyst](https://aiwiki.ai/wiki/general_catalyst), [Kleiner Perkins](https://aiwiki.ai/wiki/kleiner_perkins) and [Sequoia Capital](https://aiwiki.ai/wiki/sequoia_capital), and named [Microsoft](https://aiwiki.ai/wiki/microsoft), Notion and [Harvey](https://aiwiki.ai/wiki/harvey_ai) as early partners.[1] Calcalist reported that the round valued the company at $600 million.[3]

The company takes its name from the neuroscience term "engram," the trace of a memory in the brain.[1][2] It is a separate subject from the "Engram" conditional-memory component that DeepSeek uses in models such as [DeepSeek V4.1-Flash](https://aiwiki.ai/wiki/deepseek_v4_1_flash).

## Overview

| Field | Value |
| --- | --- |
| Type | Private company |
| Founded | October 2025[2][3] |
| Headquarters | San Francisco, California[1] |
| Founders | Dan Biderman, Sabri Eyuboglu, Jessy Lin, Jack Morris, Scott Linderman, Christopher Ré[1] |
| Key people | Dan Biderman (CEO), Sabri Eyuboglu (CTO)[1] |
| Focus | Memory and continual learning for language models and agents[4][8] |
| First product | An API for agents that learn on large shared knowledge workspaces[4] |
| Employees | 13 (June 2026)[2][6] |
| Funding | $98 million (June 2026)[1] |
| Reported valuation | $600 million (June 2026, per Calcalist)[3] |
| Website | engram.com |

## History

### Founding

According to General Catalyst, Biderman and Eyuboglu met in Christopher Ré's research lab at [Stanford University](https://aiwiki.ai/wiki/stanford_university) and found they had been "chasing the same unfashionable idea from opposite ends." Biderman came to AI from theoretical neuroscience, while Eyuboglu had spent his PhD on the memory mechanisms inside transformers. The team grew to include Lin, Morris and Linderman as co-founders.[5] Biderman told Calcalist that the company was founded "last October directly out of Stanford University's AI lab," and CNBC described Engram at its June 2026 launch as founded in October and eight months old.[2][3] In a later podcast interview, Biderman said that much of the company "started with us taking a deep dive into KV caches and caching."[8]

Biderman told CNBC that his interest in memory began in childhood, when he tried to get his grandmother, who had lost her memory, to remember small facts about him and his siblings.[2] He told Calcalist that when the founders started building the company, token costs were not a major concern in the industry and investors backed them mainly for their technical and academic record. In his account, token consumption rose sharply at the start of 2026 as new model capabilities arrived, and cost became a pressing issue for organizations.[3]

### Launch

Engram announced its $98 million funding and came out of stealth on June 23, 2026.[1] CNBC and Quartz reported that the company had 13 employees and planned to spend the money on compute and hiring.[2][6] The launch post on Engram's site said the company was already training models on its own GitHub, Slack and Notion data, ran this process on all of its company data every day, and was "moving towards retraining every hour, and eventually, every minute."[4] On June 24, 2026, the day after the launch, Biderman and Lin appeared on episode 90 of Sequoia Capital's Training Data podcast, where host Sonya Huang introduced Engram as "a NeoLab focused on memory and continual learning."[8]

### Legal research with Harvey

On August 17, 2026, Engram published results from a study built with Harvey on a synthetic law firm, described below.[9] Harvey described the same work three days later in the research preview for its [Harvey Tenet](https://aiwiki.ai/wiki/harvey_tenet) legal model, as one of several capabilities that it trained separately and could deploy as tools or sub-agents.[10]

### Later activity

As of late October 2026, the company's careers page listed openings for research scientists and for platform, product, ML systems and security infrastructure engineers.[23]

## Founders

| Name | Role | Background |
| --- | --- | --- |
| Dan Biderman | Co-founder, CEO | PhD at Columbia University's Center for Theoretical Neuroscience; postdoctoral work at Stanford under Christopher Ré on making AI agents cheaper to run.[1] Lead author of "LoRA Learns Less and Forgets Less" (2024), written with [Databricks](https://aiwiki.ai/wiki/databricks) Mosaic Research.[14] Calcalist describes him as an Israeli researcher.[3] |
| Sabri Eyuboglu | Co-founder, CTO | Stanford PhD under Christopher Ré; lead author of Cartridges (2025).[1][11] |
| Jessy Lin | Co-founder | Berkeley PhD who did research at Meta's [FAIR](https://aiwiki.ai/wiki/meta_fair) lab; developed Active Reading.[1][12] Her site lists Anca Dragan and Dan Klein as PhD advisers and an MIT undergraduate degree in computer science and philosophy.[19] |
| Jack Morris | Co-founder | Cornell PhD (Cornell Tech) who worked part time as a student researcher at Meta ([FAIR](https://aiwiki.ai/wiki/meta_fair)) and was earlier a Google Brain Resident.[20] Known for work on retrieval and on memorization in language models.[1][15] |
| Scott Linderman | Co-founder | Associate professor of statistics at Stanford and member of the Wu Tsai Neurosciences Institute; works on [state space models](https://aiwiki.ai/wiki/state_space_model) and computational neuroscience.[21] |
| Christopher Ré | Co-founder | Professor of computer science at Stanford, in the Stanford AI Lab.[22] Supervised Biderman's postdoctoral work and Eyuboglu's PhD.[1] |

## Research background

Engram says its members have worked on memory "from every angle," naming context compression, retrieval, [LoRA](https://aiwiki.ai/wiki/lora), [synthetic data](https://aiwiki.ai/wiki/synthetic_data), long-context and memory architectures, and the study of memorization and forgetting in humans and machines.[4] General Catalyst wrote that the company's approach "grew out of" a run of the founders' research, listing BASED, Cartridges, "LoRA fundamentals," Minions, Lin's Active Reading and sparse memory finetuning, and Morris's memorization work.[5] VKTR described the technology as rooted in the June 2025 Cartridges paper.[7]

| Paper | Year | Founders among authors | Main result reported by the authors |
| --- | --- | --- | --- |
| Simplified State Space Layers for Sequence Modeling (S5) | 2022 | Linderman | Replaced the many single-input, single-output state space models of an S4 layer with one multi-input, multi-output model that can use parallel scans.[18] |
| Simple linear attention language models balance the recall-throughput tradeoff (BASED) | 2024 | Eyuboglu, Ré | Identified a tradeoff between a model's recurrent state size and its recall ability, and proposed BASED, which combines linear and sliding-window attention.[17] |
| LoRA Learns Less and Forgets Less | 2024 | Biderman | On code and math, low-rank LoRA underperformed full [fine-tuning](https://aiwiki.ai/wiki/fine_tuning) but better preserved the base model's performance outside the target domain.[14] |
| Minions | 2025 | Biderman, Eyuboglu, Linderman, Ré | A small on-device model working with a cloud frontier model; the MinionS protocol cut costs by 5.7 times on average while recovering 97.9% of the frontier model's performance.[16] |
| How much do language models memorize? | 2025 | Morris | Estimated that GPT-style models have a capacity of about 3.6 bits per parameter.[15] |
| Cartridges | 2025 | Eyuboglu, Ré | Trained a small [KV cache](https://aiwiki.ai/wiki/kv_cache) offline for each corpus using "self-study" on synthetic conversations; on long-context benchmarks it matched [in-context learning](https://aiwiki.ai/wiki/in_context_learning) while using 38.6 times less memory and enabling 26.4 times higher throughput.[11] |
| Learning Facts at Scale with Active Reading | 2025 | Lin | Trained models to study material with self-generated learning strategies; an 8B expert model reached 66% on a Wikipedia-grounded subset of SimpleQA.[12] |
| Continual Learning via Sparse Memory Finetuning | 2025 | Lin | Updating only the memory slots most activated by new knowledge reduced forgetting: NaturalQuestions F1 fell 11%, against 89% after full fine-tuning and 71% with LoRA.[13] |

Four of the six co-founders (Biderman, Eyuboglu, Linderman and Ré) were authors of the Minions paper, published in February 2025.[16]

## Technology

### Learned memory

Engram's launch materials describe models that "study an organization's world and anticipate its questions in advance," forming a compact memory unique to each customer that keeps improving as it is used.[1] The company contrasts this with frontier-lab training: rather than spending training compute on public data, it starts from strong pre-trained models and spends compute on the customer's own context.[4] It describes its long-term goal in terms of [continual learning](https://aiwiki.ai/wiki/continual_learning): "a single training algorithm that can absorb arbitrary amounts of data into a model that gets continually better."[4] Eyuboglu explained the cost argument in the launch release:

> "When an AI reads a 70,000-word legal contract, which is roughly 400 kilobytes of text, its internal memory of that document can swell past 100 gigabytes. That's 250,000 times larger than the original file, and a huge part of what makes AI slow and expensive to run," said Eyuboglu. "We do that studying once, ahead of time, training the model to compress everything it learns into a compact memory it can reuse on every query."[1]

Calcalist described the architecture as separating a model's "reasoning and inference" layer from its "memory" layer, so that models can adapt to user preferences, conversation history and new data within seconds or hours.[3]

### Training methods

In the Sequoia interview, Lin said Engram trains per-team models inside customer workspaces using adapter fine-tuning "of many types," such as LoRAs, prefixes or sparse architectures, and turns raw documents and interactions into training signal with supervised fine-tuning, [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) and on-policy distillation.[8] Biderman said the method needs white-box access to model weights. He said Engram can work with companies that have closed weights, but it is easiest with open-source models, and the technique applies to any transformer model.[8] The founders said the approach does not replace tools such as search: the models assume some knowledge stays external, and the open problem is deciding what to internalize. Biderman called it "a RAG killer kind of thing" while adding that the models can still use [retrieval-augmented generation](https://aiwiki.ai/wiki/retrieval_augmented_generation) and other retrieval.[8]

Engram's August 2026 research post said it trains its agents on open-source base models with parameter-efficient fine-tuning. It said it was interested in combining LoRA, trainable KV caches and added experts in [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) models.[9] In that system the agent keeps two kinds of memory: text notes, which are easy to inspect and edit, and parametric memory in the trained weights, which captures diffuse associations that are hard to write down. The agent is trained end to end to use both alongside search.[9]

### Efficiency claims

Engram's claims about cost come from the company and its partners. The launch release said its models match or outperform frontier models while using 1 to 10% of the tokens.[1] Biderman told Sequoia that a model trained on a team's data can "answer within 100 tokens what the best frontier models would consume 100,000 tokens doing" for questions about people, teams and priorities that no single document answers.[8] He also told CNBC that Engram's models are not "absolutely better" than those of OpenAI or Anthropic. In his account they excel at specializing, sometimes at the expense of other capabilities.[2]

## Partnerships

Engram launched with three named partners. CNBC described them as clients.[1][2]

| Partner | Work described | Sources |
| --- | --- | --- |
| Microsoft | Testing Engram models inside Microsoft 365 to build cost-efficient, customized agents for enterprise customers. The partnership includes a commitment of GPU capacity, including on [Microsoft Azure](https://aiwiki.ai/wiki/azure), for training Engram's models. | [1][4] |
| Notion | Design partner. Engram is building custom agents that understand large Notion workspaces, and Notion is testing Engram's models inside its new custom agents. | [1][4] |
| Harvey | Developing models that internalize a whole firm's knowledge and can find precedents across many client matters. | [1][4][10] |

Jason Graefe, corporate vice president for AI Partner Catalyst at Microsoft, called the work "the sort of frontier bet we want to be making."[1] Notion co-founder Simon Last said that in testing, Engram's models were approaching frontier quality "while using an order of magnitude fewer tokens, because the agent already knows the workspace instead of rediscovering it on every query."[1] Harvey co-founder and president Gabe Pereyra said the two companies were building "learned enterprise memories."[1]

## Calderwood and Harkness study

Engram's first published research result, released on August 17, 2026 "in partnership with Harvey," tested an agent on a synthetic law firm called Calderwood & Harkness (C&H).[9] The firm's filesystem holds 266 matter folders, 9,286 files and about 100 million tokens of emails, memos, legal documents and spreadsheets. Tasks are precedent-search and reasoning questions that depend on the firm's history, such as finding the firm's most recent M&A deal with a working-capital adjustment.[9] Harvey refers to the benchmark as LAB: Firm Knowledge.[10]

During training, the agent explored the workspace over thousands of rollouts, wrote structured notes that compressed the 100 million tokens into about 1 million tokens, and trained lightweight knowledge modules through distillation on model outputs and reinforcement learning on problems the model generated itself.[9] The base model was the dense [Qwen3.8](https://aiwiki.ai/wiki/qwen3_8)-27B.[9]

| Measure (Engram's reported results) | Engram agent | Comparison |
| --- | --- | --- |
| Average cost per query | $0.13 | $1.32 for [Claude Opus 4.8](https://aiwiki.ai/wiki/claude_opus_4_8) at high thinking level |
| All-pass rate | 30% | 25% for Claude Opus 4.8 |
| Closed-book accuracy on questions about the firm | 72.6% after studying | 4.7% before studying |
| First turn begins with a targeted search | 99.2% of trajectories | 3.2% for the base Qwen3.8-27B |

Engram computed costs from [OpenRouter](https://aiwiki.ai/wiki/openrouter) prices as of August 16, 2026, assuming perfect caching across agent turns, and reported that spending more training compute on studying pushed cost per query down further.[9] Harvey's own summary said the trained model improved criteria pass rate by more than 15% and task completion rate by nearly 10%, cut total tokens in completed trajectories by 58%, and reduced cost per query by 90% while completing tasks competitively with leading frontier models.[10] Both accounts come from the companies that did the work, and the law firm is synthetic. Engram described its compute-scaling results as "only initial promising signs."[9]

## Funding

| Date | Amount | Investors | Valuation |
| --- | --- | --- | --- |
| Announced June 23, 2026 | $98 million | General Catalyst, Kleiner Perkins, Sequoia Capital, Factory, Modern, Amplify Partners, Neo, SV Angel and others; angels and advisers including Assaf Rappaport, [Andrej Karpathy](https://aiwiki.ai/wiki/andrej_karpathy) and [Pieter Abbeel](https://aiwiki.ai/wiki/pieter_abbeel)[1][4] | $600 million, as reported by Calcalist; the company did not state a valuation in its release[1][3] |

General Catalyst wrote that it co-led the company's first round of funding with Modern Capital.[5] Quartz described the round as led by General Catalyst, Kleiner Perkins and Sequoia Capital, and Calcalist said those three firms were among the leaders.[3][6] Calcalist also reported that the private investors included Merav Bahat and Tomer Schwartz, Ofir Ehrlich, Roi Tiger and Ilan Twig, and that Rappaport led a group of Israeli technology entrepreneurs backing the company.[3]

Engram has not disclosed revenue. Biderman told Calcalist in June 2026 that the company was "already on track to generate significant revenue through contracts signed with some of the largest companies in the market."[3]

## Reception

Investors presented the company as an answer to rising inference costs. General Catalyst described today's models as "brilliant and amnesic at once" and said the fix was to train knowledge into model weights instead of re-feeding the same material at inference time.[5] General Catalyst chief executive Hemant Taneja called memory "the missing ingredient in AI," and Kleiner Perkins partner Leigh Marie Braswell told CNBC that Engram "basically maps out your organization and offers orders of magnitude cheaper output."[1][2] CNBC and Quartz placed the launch alongside a broader push by large companies to cap or scrutinize AI spending.[2][6]

The token-efficiency figures behind that pitch are reported by Engram and its partners rather than by independent evaluators, and Biderman has said that the models' specialization sometimes comes at the expense of other capabilities.[1][2][9][10] Engram is included in Debarghya Das's "The Ultimate Neolab List" of AI research companies, and Sequoia's podcast used the same label.[8][24]

## See also

- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)
- [Continual learning](https://aiwiki.ai/wiki/continual_learning)
- [KV cache](https://aiwiki.ai/wiki/kv_cache)
- [Retrieval-augmented generation](https://aiwiki.ai/wiki/retrieval_augmented_generation)
- [Context engineering](https://aiwiki.ai/wiki/context_engineering)
- [Harvey Tenet](https://aiwiki.ai/wiki/harvey_tenet)
- [Notion AI](https://aiwiki.ai/wiki/notion_ai)

## References

1. Engram (via PR Newswire). ["Engram Launches with $98M to Build AI That Actually Knows Your Organization"](https://www.prnewswire.com/news-releases/engram-launches-with-98m-to-build-ai-that-actually-knows-your-organization-302807126.html). June 23, 2026.
2. Samantha Subin, CNBC. ["AI memory startup focused on cutting token costs raises $98 million"](https://www.cnbc.com/2026/06/23/ai-memory-startup-focused-on-cutting-token-costs-raises-98-million.html). June 23, 2026.
3. Meir Orbach, Calcalist (Ctech). ["AI startup Engram raises $98 million at $600 million valuation with just 13 employees"](https://www.calcalistech.com/ctechnews/article/skoqbrumze). June 23, 2026.
4. Engram. ["Introducing Engram"](https://engram.com/blog/introducing-engram). June 23, 2026.
5. Marc Bhargava and Max Rimpel, General Catalyst. ["Our Investment in Engram"](https://www.generalcatalyst.com/stories/our-investment-in-engram). June 23, 2026.
6. Colleen Cabili, Quartz. ["AI memory startup Engram raised $98 million to help companies spend less running enterprise AI"](https://qz.com/engram-ai-memory-startup-98-million-funding-062426). Updated July 3, 2026.
7. Sheryl Hodge, VKTR. ["Engram Launches With $98M to Give Enterprise AI Memory"](https://www.vktr.com/ai-news/engram-launches-with-98m-to-give-enterprise-ai-memory/). June 24, 2026.
8. Sequoia Capital. ["Memory and Continual Learning: Engram's Dan Biderman and Jessy Lin"](https://sequoiacap.com/podcast/memory-and-continual-learning-engrams-dan-biderman-and-jessy-lin). Training Data, episode 90, interview transcript. June 24, 2026.
9. Engram. ["Understanding a Law Firm through Study"](https://engram.com/blog/legal-agents-with-memory). August 17, 2026.
10. Calvin Qi, Vasudha Rengarajan, Julio Pereyra, Niko Grupen and Gabe Pereyra, Harvey. ["Harvey Tenet Research Preview"](https://www.harvey.ai/blog/post-training-update-harvey-tenet). August 20, 2026.
11. Sabri Eyuboglu, Ryan Ehrlich, Simran Arora, et al. ["Cartridges: Lightweight and general-purpose long context representations via self-study"](https://arxiv.org/abs/2506.06266). arXiv:2506.06266, June 6, 2025.
12. Jessy Lin, Vincent-Pierre Berges, Xilun Chen, Wen-tau Yih, Gargi Ghosh and Barlas Oğuz. ["Learning Facts at Scale with Active Reading"](https://arxiv.org/abs/2508.09494). arXiv:2508.09494, August 13, 2025.
13. Jessy Lin, Luke Zettlemoyer, Gargi Ghosh, et al. ["Continual Learning via Sparse Memory Finetuning"](https://arxiv.org/abs/2510.15103). arXiv:2510.15103, October 16, 2025.
14. Dan Biderman, Jacob Portes, Jose Javier Gonzalez Ortiz, et al. ["LoRA Learns Less and Forgets Less"](https://arxiv.org/abs/2405.09673). arXiv:2405.09673, May 15, 2024.
15. John X. Morris, Chawin Sitawarin, Chuan Guo, et al. ["How much do language models memorize?"](https://arxiv.org/abs/2505.24832). arXiv:2505.24832, May 30, 2025.
16. Avanika Narayan, Dan Biderman, Sabri Eyuboglu, Avner May, Scott Linderman, James Zou and Christopher Ré. ["Minions: Cost-efficient Collaboration Between On-device and Cloud Language Models"](https://arxiv.org/abs/2502.15964). arXiv:2502.15964, February 21, 2025.
17. Simran Arora, Sabri Eyuboglu, Michael Zhang, et al. ["Simple linear attention language models balance the recall-throughput tradeoff"](https://arxiv.org/abs/2402.18668). arXiv:2402.18668, February 28, 2024.
18. Jimmy T. H. Smith, Andrew Warrington and Scott W. Linderman. ["Simplified State Space Layers for Sequence Modeling"](https://arxiv.org/abs/2208.04933). arXiv:2208.04933, August 9, 2022.
19. Jessy Lin. ["Jessy Lin"](https://jessylin.com/) (personal website). Accessed October 27, 2026.
20. Jack Morris. ["About"](https://jxmo.io/) (personal website). Accessed October 27, 2026.
21. Stanford Profiles. ["Scott Linderman"](https://profiles.stanford.edu/scott-linderman). Accessed October 27, 2026.
22. Stanford Profiles. ["Christopher Ré"](https://profiles.stanford.edu/christopher-re). Accessed October 27, 2026.
23. Engram. ["Careers"](https://engram.com/careers). Accessed October 27, 2026.
24. Debarghya Das. ["The Ultimate Neolab List"](https://debarghyadas.com/neolabs/). Accessed October 27, 2026.