Engram
by Engram
AI models that study your organization's knowledge before they answer
Engram trains a compact, reusable memory of an organization's documents, code and communications ahead of time, so agents answer from institutional knowledge without re-reading the same context on every query. Its first product is an API for agents that work inside large shared knowledge workspaces, where repeated context gathering drives token spend.
Engram is a San Francisco startup that builds AI models which learn an organization's own context before they are asked anything, instead of retrieving and re-reading documents on every request. The company emerged from stealth on June 23, 2026 with a $98 million Series A from General Catalyst, Kleiner Perkins and Sequoia Capital, with Factory, Modern, Amplify Partners and Neo participating and angels including Andrej Karpathy, Assaf Rappaport and Pieter Abbeel. Its founders come out of Stanford, Berkeley, Cornell and Meta FAIR research: CEO Dan Biderman, CTO Sabri Eyuboglu, Jessy Lin, Jack Morris, Scott Linderman and Chris Re. The technical basis is research the team published before founding the company. Cartridges, a 2025 paper led by Eyuboglu, trains a small KV cache offline on a corpus through 'self-study' on synthetic conversations, and reported matching in-context learning while using 38.6x less memory with 26.4x higher throughput. Lin's Active Reading work covers how a model studies a corpus in depth. Engram's pitch is that spending training compute on the customer's context, starting from strong pretrained models, makes agents 10x to 100x more token-efficient on repeated work. The first product is an agent API for large shared knowledge workspaces. Named partners at launch: Notion (Custom Agents for large workspaces), Harvey (models that internalize a law firm's matters) and Microsoft, which is piloting Engram models inside Microsoft 365. Press coverage put the team at about 13 people at launch, so for now this is an early, partner-led platform, not a self-serve product.
A platform or AI engineering lead running agents over a large internal corpus, where token spend from re-retrieving the same context is already a budget line
Agents that answer from pre-trained organizational memory, which the vendor says needs 10x to 100x fewer tokens per task
At a Glance
- Category
- AI Models & APIs
- Pricing
- Contact for pricing
- Target Market
- CTOs, CIOs, AI Platform Teams, Enterprise Developers
- Deployment
- API-based
- Founded
- 2025
- Headquarters
- San Francisco, USA
- Team Size
- 11-50
Key Features
- ✓Learned organizational memory
Trains a compact, reusable memory of company documents, code and communications ahead of time, so each query does not start from zero.
- ✓Cartridges
Offline-trained KV caches built by self-study on synthetic conversations; the research paper reported 38.6x less memory than in-context learning.
- ✓Active Reading
A study phase in which the model works through the corpus in depth before query time, building knowledge it later answers from.
- ✓Agent API for shared workspaces
The first commercial product: an API that lets agents draw on a large shared knowledge base without gathering the same context repeatedly.
- ✓Hybrid memory (weights, notes, search)
In the Harvey legal study the agent combined parametric knowledge, compressed text notes and targeted search over a 100M-token workspace.
- ✓Customer-owned memories
The company positions trained memories as belonging to the customer, so the model improves the longer an organization uses it.
Capabilities
Use Cases
- •Legal precedent search
With Harvey, an agent studied a synthetic 100M-token law firm and answered at $0.13 per query against $1.32 for a frontier model.
- •Workspace agents in Notion
Notion is building Custom Agents for large workspaces on Engram, so agents know the workspace without re-reading it each time.
- •Microsoft 365 enterprise agents
Microsoft is piloting Engram models in Microsoft 365 to build agents customized to an individual enterprise customer's knowledge.
- •Cutting repeated context costs
Teams whose agents re-retrieve the same documents across many tasks can move that cost into one up-front training step.
Ideal For
Best For
- ✓Agents that repeatedly query the same large shared workspace, such as a wiki or document store
- ✓Legal teams searching precedent across a firm's own client matters
- ✓Software vendors embedding organization-aware agents in their product, as Notion is doing
- ✓Enterprises whose agent token bills are dominated by context gathering
Not Ideal For
- ✗Teams that need a self-serve, published-price product today; access runs through the company and partner pilots
- ✗Workloads over fast-changing data, since the company itself calls continual learning on edited documents an open research problem
- ✗Buyers who need documented compliance certifications before a pilot; none are published on the vendor site
Deployment
Market Analysis
Pros
- ✓Peer-reviewable research basis: the Cartridges paper reports 38.6x less memory and 26.4x higher throughput than in-context learning
- ✓Large funding and notable launch partners for a company of its size
- ✓Published a concrete cost comparison in the Harvey study ($0.13 vs $1.32 per query)
Cons
- ✗Very early: about 13 employees at launch and no self-serve access or public pricing
- ✗Its own blog calls continual learning on changing documents a research problem with 'much work to be done'
- ✗The Harvey results came from a synthetic law firm, and the post discloses no production deployment, security or compliance detail
- ✗Headline 100x token savings are vendor claims, not independently benchmarked
Pricing
Enterprise
Contact for pricing
- ✓Agent API
- ✓Memory trained on the customer's own corpus
- ✓Access via the company (hello@engram.com)
No list pricing is published. The site offers a contact address only, and the launch announcement describes partner deployments, so expect a negotiated enterprise engagement. The cost model trades per-query tokens for up-front training compute on your corpus.
Security & Compliance
Sources
This page was written from 6 sources, 4 on domains other than engram.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Arena AI Evaluations
Paid model evaluations built on the Arena leaderboard's crowdsourced human votes, for enterprises choosing an LLM
Kimi K3
2.8T-parameter open-weight MoE model with a 1M-token context, priced at $3/$15 per million tokens
TypeSafe Jev
A 'System One' decision model that returns typed choices, scores and calibrated probabilities instead of text, for agent routing and classification
Xiaomi MiMo-V2.6
MIT-licensed, 1M-context multimodal MoE models that took the top open-weights score on Artificial Analysis at well under $1 per million tokens