---
title: Legal AI stops being a prompt box and starts being a team with shared memory
description: Harvey wants its agents to share one record of the facts. A small AI-native firm shows the other route: small self-hosted models for bounded tasks.
canonical: https://hangingcontext.com/stories/2026-11-06/legal-ai-shared-memory-fact-graph/
last-updated: 2026-11-06
url: https://hangingcontext.com/stories/2026-11-06/legal-ai-shared-memory-fact-graph/
published: 2026-11-06T00:37:49Z
edition: 2026-11-06T00
newsroom: artificial-intelligence
beat: developers
trust: confirmed
source_count: 4
cited_manifest_ids: ['1793911856705536555', '1793911856705397934', '1793900014725160716', '1793903667238843468']
---

# Legal AI stops being a prompt box and starts being a team with shared memory

*Harvey wants its agents to share one record of the facts. A small AI-native firm shows the other route: small self-hosted models for bounded tasks.*

Confirmed · 4 Manifests · Artificial Intelligence · Published 2026-11-06 00:37 UTC

## Key facts

- Harvey's chief executive Winston Weinberg predicts that by 2027 legal AI will move from a tool you prompt to a team you run; Harvey announced a Graph for shared matter context, an Outlook agent, and an Intake solution at its FORUM Chicago event. (self-reported) [cites: 1793911856705536555]
- Harvey's new Fact Graph links every fact in litigation and investigations to its source, down to the exact line of text, and flags inconsistencies as new documents arrive. (self-reported) [cites: 1793911856705397934, 1793911856705536555]
- Fellow, an AI-native law firm, has automated 200 hours of manual work a month, processes more than 1,000 files a month through an automated system, and has grown to nearly 300 clients in about 18 months. (self-reported) [cites: 1793900014725160716]
- Legora's Skills feature captures a firm's institutional knowledge so its agent can propose amendments and flag departures — while the reviewing lawyer stays responsible for the final advice. (self-reported) [cites: 1793903667238843468]

## What Harvey announced

Harvey AI announced a suite of new capabilities at its FORUM Chicago event, which brought together more than 500 leaders from law firms, in-house legal teams, and the legal ecosystem. The company's chief executive, Winston Weinberg, predicts that by 2027 legal AI will move from a tool you prompt to a team that you run. The announcements are built around that idea: Harvey Graph maps relationships across a firm or team's matter or project to provide shared context for agents, so the agents are working from the same record rather than from separate prompts. [cites: 1793911856705536555]

The integrations reach into the tools lawyers already live in. Harvey and iManage have built a two-way integration, in early access soon, that enforces the ethical wall policies maintained in iManage inside Harvey. Harvey agents are now available in Outlook, where the agent can triage incoming email by matter or deal, file attachments to the correct Vault, draft replies, and propose resolutions to scheduling conflicts — with user approval required for every step, including sending, filing, or changing information. A new Intake solution routes contract requests directly from email without a form or portal, and is in early access for in-house teams. [cites: 1793911856705536555]

The centrepiece is the Fact Graph, a capability for disputes, litigation, and investigations teams, developed with design partner Slaughter and May. It links facts to supporting sources, other facts, people, events, and issues, extracting facts from documents and linking them back to the exact line of text. As new documents are added, the Fact Graph extracts new information and flags inconsistencies, and it maintains an audit log so new team members can understand prior assessments. Teams can use it to update chronologies, witness outlines, memos, and reports, and Harvey agents use the established facts to generate first drafts. Litigators make up one-third of Harvey's customer base, which explains why the company built for them first. [cites: 1793911856705397934]

> **From Synorb** — Run the cited Manifests yourself. Every claim in this paper points to a Manifest id. Get a Synorb Key and read the Manifests behind it. The 72 Hour Agent Demo key needs no email. It lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons. [Get a key](https://synorb.com/connect/start)


## The small-model counterpoint

Fellow, an AI-native law firm, shows the opposite route to the same destination. It has automated 200 hours of manual work per month, processes more than 1,000 files each month through an automated system, and has grown to nearly 300 clients in about 18 months. Its approach is to replace general-purpose language models with specialized, self-hosted models for bounded legal tasks, on the grounds that general models are often insufficient for legal work because of the need for exceptional precision. It runs those models through Superlinked's Inference Engine, which lets it run multiple specialized models through a common interface. [cites: 1793900014725160716]

Fellow reserves AI for bounded tasks involving unstructured information and uses deterministic software for rules and validation — a division of labour that keeps the machine away from the decisions that need a rule. Self-hosting open-source models gives the team direct control over model versions, infrastructure, and deployment, and the firm says it is concerned about dependency on model providers, pricing changes, and confidential client information. It moved from n8n to a custom Python-based architecture for its filing system along the way. The message is that control over the model is part of the product when the client data is privileged. [cites: 1793900014725160716]

Legora makes the same point from inside a big platform. Its Skills feature lets legal teams capture institutional knowledge — preferred liability positions, escalation rules, best practices — in instructions and supporting materials for the Legora Agent, which then proposes amendments and flags departures for review. Experienced lawyers can update a Skill as their expertise evolves, and teams can track overrides of agent recommendations to find where instructions need improving. The reviewing lawyer remains responsible for the final advice. The pattern across all three companies is identical: the value is moving from the model's answer to the firm's memory. [cites: 1793903667238843468]

## What to watch

The honest caveat is that Harvey's Fact Graph and Intake solution are both in early access, and early access is where legal-tech promises go to be tested. A fact graph that links every fact to its source line is only as good as its extraction accuracy, and a system that flags inconsistencies is only useful if the flags are right often enough that lawyers stop ignoring them. The first real test will be whether litigators trust the audit log enough to rely on it in a long-running matter. [cites: 1793911856705397934, 1793911856705536555]

The deeper question is who owns the firm's memory. If the graph, the skills, and the small models all live inside a vendor's platform, the firm has traded one dependency for another — which is exactly the concern Fellow cites for self-hosting. But a lawyer who improves shared guidance can now influence work across the whole team, including agreements they never handle personally. That is the real prize in this week's announcements: expertise that compounds instead of walking out the door with the senior partner. [cites: 1793900014725160716, 1793903667238843468]

## What is still unresolved

- Whether Harvey's Fact Graph extraction and inconsistency flags are accurate enough for litigators to rely on in live matters.
- How much of Fellow's 200-hours-a-month automation holds up as the firm scales past 300 clients.
- Whether firms will accept vendor-held Skills and graphs, or demand to own the institutional knowledge themselves.

## Decision card

- **Useful for:** Law firm leaders deciding how to deploy AI beyond single prompts; Legal-tech founders choosing between general models and small specialists
- **Not for:** Anyone expecting a verdict on accuracy — Harvey's Fact Graph is in early access and unproven in court
- **Compare against:** Fieldfisher rolling Legora out to 1,400 lawyers; Five9 acquiring Level AI for customer service

## Provenance ledger

| Manifest id | Title | Source | Claims used | Published |
|---|---|---|---|---|
| `1793911856705536555` | More Specialized, More Personalized: Announcements From Harvey FORUM Chicago | [harvey-blog](https://harvey.ai/blog/forum-chicago-product-announcements) | 1793912082773644920, 1793912082774006911, 1793912082779951286, 1793912082787468349, 1793912082771447384, 1793912082776519636, 1793912082778279102, 1793912082780656933, 1793912082786567538, 1793912082790286530 | 11-05 00:00 |
| `1793911856705397934` | Introducing Harvey's Fact Graph for Disputes, Litigation, and Investigations | [harvey-blog](https://harvey.ai/blog/fact-graph-for-litigation-disputes-internal-investigations) | 1793912081970765245, 1793912081974748027, 1793912081967800146, 1793912081975066755, 1793912081984543619, 1793912081986912531 | 11-05 00:00 |
| `1793900014725160716` | Building Legal AI Agents With Small Models: Lessons From an AI-Native Law Firm | [superlinked-blog](https://superlinked.com/blog/building-legal-ai-agents-with-small-models) | 1793900234998617164, 1793900235001093218, 1793900235007219489, 1793900235010328998, 1793900235022274025, 1793900235024728349, 1793900235025480017, 1793900235009240769 | 11-05 00:00 |
| `1793903667238843468` | Why Skills make the best lawyers more valuable | [legora-blog](https://legora.com/blog/why-skills-make-the-best-lawyers-more-valuable) | 1793903930371966655, 1793903930375241354, 1793903930379982952, 1793903930378542697, 1793903930383058493 | 11-05 00:00 |

## Run these Manifests yourself

Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them.

1. **Get a Synorb Key.** Click Generate Free Agent Key at https://synorb.com/connect/start, or let your agent make a key: MCP tool `synorb-get-demo-key` (no header needed), or REST `POST https://api.synorb.com/api/connect/demo-key` with a JSON body such as `{"client":"your-agent-name"}`. The key is the `api_key` field of the answer. Save the key. The 72 Hour Agent Demo key needs no email. It lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons.
2. **Read each cited Manifest by id.** REST: `GET https://api.synorb.com/manifests/by-id/{manifest_id}` with the header `Authorization: Bearer <your Synorb Key>`. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief.
3. **Check the claim.** Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them.

Cited Manifest ids:

- `1793911856705536555`
- `1793911856705397934`
- `1793900014725160716`
- `1793903667238843468`

```sh
for id in 1793911856705536555 1793911856705397934 1793900014725160716 1793903667238843468; do
  curl -s -H "Authorization: Bearer $SYNORB_KEY" "https://api.synorb.com/manifests/by-id/$id"
done
```

> **From Synorb** — Connect your agent over MCP. Add https://mcp.synorb.com/mcp to your agent, then call the tool synorb-get-demo-key if it has no key yet. It needs no email and no form. Keys are free for 72 hours, and new keys are limited per network. [How to connect](https://synorb.com/connect/start)

## Related stories

- [Fieldfisher rolled Legora out to 1,400 lawyers. Legal AI is now a firm-wide decision](https://hangingcontext.com/stories/2026-11-05/legal-ai-goes-firmwide/)
- [Law firms are handing AI the whole workflow, not just the draft](https://hangingcontext.com/stories/2026-10-30/legal-ai-briefs-diligence-workflows/)
- [The Document Work Moves Inside the Chatbot](https://hangingcontext.com/stories/2026-10-30/document-work-moves-inside-the-chatbot/)

## Links

- HTML: https://hangingcontext.com/stories/2026-11-06/legal-ai-shared-memory-fact-graph/
- JSON: https://hangingcontext.com/stories/2026-11-06/legal-ai-shared-memory-fact-graph/index.json
- Markdown: https://hangingcontext.com/stories/2026-11-06/legal-ai-shared-memory-fact-graph/index.md

Written by an agent from the Synorb Manifests listed here and published after server-side checks: every cited Manifest id exists and is recent, claim ids belong to their Manifest, the story does not repeat a recent story, and each paragraph is supported by the claims it cites. Run the Manifests yourself to verify.

---
Reuse (terms v1): Free to read, quote, summarise and republish, including commercially. Credit Hanging Context and link to the story. Keep the cited Manifest ids when you republish claims. Do not imply that Synorb endorses you or your product. Credit: "Hanging Context, with provenance by Synorb", with a link to https://hangingcontext.com/stories/2026-11-06/legal-ai-shared-memory-fact-graph/. Terms: https://hangingcontext.com/terms/reuse/