A large firm's knowledge is spread across the DMS, the file server, the mailboxes and whatever the last IT project left behind. LegalMemory reads all of it where it sits and turns it into one index a lawyer, or an AI tool, can actually ask a question of.
Documents stay in iManage, SharePoint, NetDocuments, the file server and the mailbox, and those systems stay the record. LegalMemory builds an index from them instead of replacing them, so there is no migration project and nobody moving files by hand.
No migrationMost search tools will hand back a draft from two years ago. LegalMemory knows which document replaced which, so an answer quotes the signed version and says so when it is deliberately citing an older draft.
Drafts, redlines, signedEvery document sits in its matter, its practice area and its document type, with the parties and dates attached. Ask about one deal and you get that deal's file, not every share purchase agreement in the firm.
Matter, practice area, partiesNew documents, new versions and withdrawn access reach the index on their own, usually within minutes. Nobody has to remember to upload anything, and nothing goes stale between IT projects.
Minutes, not quartersThis is the part that decides whether a large firm can use a knowledge index at all. LegalMemory installs into the firm's own environment, so client files, privileged material and the index built from them never cross the perimeter.
Every component runs as a container on the firm's servers or in a tenant the firm controls. Nothing calls out. This is the reason a hosted index is a non-starter for most firms of a certain size.
On premises or your tenantAccess rights are copied from the source system and tied to the person through single sign-on. If somebody cannot open a document in iManage, it does not show up in their results, and it never influenced their ranking either.
The same walls as your DMSReading and indexing run through models the firm chooses, hosted under its own agreement or on its own hardware. No client data trains anybody else's model, and no vendor ends up holding a copy of your corpus.
No data leavesThe whole system is open source. Replace the AI product sitting on top of it next year and the index, the structure and the work that went into it stay with the firm.
Open sourceLegalMemory sits in the background and syncs against whichever DMS the firm already runs. One connector per system, authorized by the firm. Content, metadata and permissions are read; nothing is written back. The index itself does hold extracted text and embeddings, so it is a second copy of client material and has to be treated as one. That is precisely why it sits on the firm's own servers, inside the same perimeter, carrying the permissions it was built from. What goes away is the migration, the cutover date and anyone moving files by hand.
Type, practice area, matter, version status, parties and dates are read out of every document. That is how the system knows v8 is the signed version and v7 is history. Ask about the termination clause and you get the one in force, not the one that was negotiated away.
Permissions come from the source system and apply before the search, not after it. Otherwise the result list itself reveals that a matter exists, and in a firm with ethical walls that is exactly the incident nobody wants to write up.
SPA_Nordlicht_v8_execution.docxYesYesBoard_minutes_2026-02-18.pdfYesNoRedline_v7_v8.docxYesYesFee_arrangement_Nordlicht.pdfYesNoAn answer is only as good as the passage it stands on, so the honest question is whether the right one was found at all. That is what we measure, over more than 10,000 legal documents: ordinary vector search finds it in 60 out of 100 cases, LegalMemory in 95.
The figures are ours, measured on our own corpus, and they are not a public ranking. The number that will count is the one on your files, which is why a benchmark run on the firm's own corpus is part of every installation.
Harvey today, something else in two years, plus whatever the firm builds itself. They all read from the same index instead of each vendor being wired into the DMS and security-reviewed again from scratch. Every request carries the identity of the person making it.
Identity-bound tools for agents that speak the Model Context Protocol.
Search, document retrieval, and metadata for the firm's own applications.
Prebuilt links for tools that use neither MCP nor the API directly.
LegalMemory is an open-source knowledge index for law firms. It connects to the firm's existing document management, mail, and file systems, structures what it finds into matters, versions, and relationships, and serves that context to authorized AI tools over MCP or an HTTP API. It runs in the firm's own infrastructure.
No. LegalMemory runs in the firm's environment and reads source systems in place. Model calls resolve through a gateway the firm points at its own hosted or on-device models. No matter data trains a public model.
A vector index on its own returns passages that look similar. LegalMemory compiles the access scope and metadata filters first, then runs lexical, semantic, identifier, and graph retrieval inside that scope and merges the rankings. Results carry their matter, their version status, and a citation that opens the source.
They stay in the index, because version history is part of the file, but they are marked as superseded and left out of answers unless a question is explicitly about the earlier version.
Any tool that speaks MCP or HTTP. LegalWork, Harvey, Legora, Microsoft Copilot, firm-built agents, and internal software all connect the same way, and each request is bound to the identity of the person asking.
A scoping conversation, connectors for the systems that matter, a decision on the ontology, and a benchmark run on the firm's own corpus. We would rather show the number on your files than quote ours.
A conversation about the firm's systems, a connector plan, and a benchmark run on your corpus. After that a measurement decides, not a demo.