/ai

EngramLab Publishes First Research Blog on Law Firm Agents · Digg

EngramLab Publishes First Research Blog on Law Firm Agents

The post details early training of AI agents on synthetic legal data with dual memory systems.

TLDR

EngramLab released its first research blog titled Understanding a Law Firm through Study. The work shows agents trained with native memory to analyze firm records. Harvey supplied a synthetic dataset built from client matters. A 27B Qwen model trained on the data learned knowledge parametrically and through memory. Gabe Pereyra stated the model outperforms frontier models at lower cost per query. Other posts noted that memory improves personalization and reduces reliance on inference-time search tools.

Featured Source

EngramLab Publishes First Research Blog on Law Firm Agents

The post details early training of AI agents on synthetic legal data with dual memory systems.

Combined views

1.1M

17 Sources, first seen 38d ago

1.9K likes

TagsEngram Harvey Mayee Chen Jack Morris Gabe Pereyra

TagsEngram Harvey Mayee Chen Jack Morris Gabe Pereyra

38d ago

first seen 38d ago

Combined views

1.1M

17 Sources, first seen 38d ago

1.9K likes80 comments1.3K saves199 reposts

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

38d ago

first seen 38d ago

TLDR

EngramLab released its first research blog titled Understanding a Law Firm through Study. The work shows agents trained with native memory to analyze firm records. Harvey supplied a synthetic dataset built from client matters. A 27B Qwen model trained on the data learned knowledge parametrically and through memory. Gabe Pereyra stated the model outperforms frontier models at lower cost per query. Other posts noted that memory improves personalization and reduces reliance on inference-time search tools.

Featured Source

80 comments

1.3K saves

199 reposts

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

200012345678901234567890123456789

200012345678901234567890123456789

200012345678901234567890123456789

200012345678901234567890123456789

17 Sources

@EngramLabToday we're publishing our first research blog, Understanding a Law Firm through Study. We're sharing a glimpse of a future where agents are trained with native memory:

@MayeeChenRT @EngramLab: Today we're publishing our first research blog, Understanding a Law Firm through Study. We're sharing a glimpse of a future…

@scott_lindermanThis is a good time to share that I’m taking a sabbatical to work at Engram full time! Our blog post highlights some of the directions we’re taking, building agents that study and learn with complementary memory systems. There's lots of research still to be done, and I’m excited to help lead these efforts!

@EyubogluSabriWe had Qwen 3.8 study all of the docs in Harvey’s synthetic law firm - the more it studies, the fewer tokens it needs at inference time to respond to the queries

@jxmnopi'm quite excited about this! • it's cool in general that models can generate their own training data and learn from it. wasn't fully clear to me even a year ago that this would work reliably. • the thing that we're trying to build seems really important and no one has built it before • some traces from our model genuinely surprise me (e.g. the attached example, where it perfectly simulates the output of a complicated bash query, despite not being trained to do this) my main feeling overall is that there is so much to do. we see signs things are starting to work: models use their memories to produce better answers, knowing things saves lots of tokens, and all of this emerges with more compute but solving this memory calibration problem, on top of learning how to generate data in ways that scale nicely with compute, is going to take some time this is just a first step 🫡

@ItsJulioPereyraGreat working with the @EngramLab team to explore model memory as a way to improve personalization and effective knowledge recall We're particularly excited about how models improve trajectories through memory. Rather than just learning to use search tools better, memory enables more directed, efficient, and effective searches that result in far higher intelligence per token. Promising for a future where agents never start from scratch!

@harveyWe partnered with @EngramLab to train a model that learns from a firm's accumulated knowledge. We first built a 100M token synthetic law firm containing ~10k documents across 250+ matters and 46 clients. The model then internalized the firm's knowledge in two ways: 1) parametric knowledge from training on the firm's documents 2) 1M tokens of structured notes on key matter-level features These two forms of memory allow the model to understand and search the firm's work much more efficiently across 250 representative legal tasks. It bested every model we tested, with a 67% mean criteria score at just $0.13 per query - 10-24x cheaper than Opus 4.8 on high/xhigh reasoning. The model also exhibited interesting qualitative behavior after training: it began 99% of trajectories with a targeted search and reached the first relevant matter in 3.7 turns, versus 5.9 for the base model. It also acquired substantial knowledge of the firm itself. When asked questions about individual matters without using any tools, its accuracy increased from 4.7% to 72.6% over the course of training. More with @EngramLab:

@nikogrupenReally interesting result here from @EngramLab on @harvey's LAB Firm Knowledge dataset. They introduce studying as a learning mechanism that both improves quality and reduces the need for extensive inference-time reasoning, leading to Pareto efficient search over really large knowledge bases. Very bullish on leveraging upfront / offline compute to reduce online token use! Awesome to collaborate with @dan_biderman @realJessyLin and the Engram team on building more efficient long-horizon agents.

@gabepereyraAwesome research by @engramlab using the synthetic law firm we built. They trained a 27B Qwen model on synthetic client matters and used the learned knowledge to do online search more efficiently. The trained model outperforms frontier models at 10x lower cost per query. The results are promising for scaling up enterprise search and personalization.

@LM_BraswellRT @dan_biderman: Like real people, Engram's models use both neural memories and written notes to tackle hard tasks. Many legal tasks requ…

Related

Harvey and EngramLab Open Source Synthetic Law Firm

The dataset models a persistent legal work environment for testing AI agents.

Thomson Reuters Builds Custom AI Model on Qwen

The legal information company fine-tuned Alibaba's open-source model to reduce dependence on Anthropic's Claude.

VCs Highlight Open-Source Models in Legal and Finance AI

Replies note Harvey and Cfosilvia fine-tuning open-source models for legal and finance tasks.

Today's Rank

—

Not ranked yet

Today's Rank

—

Not ranked yet