Marcella

Retrieval-first legal intelligence

Matter Memory

Your practice, remembered.

Every query, citation, and analysis saved to your private tenant.Gets smarter the more you use it.

Why this is the hard part

You cannot ask the same question twice.

Run a prompt through any generative system, then run it again. The second answer will not match the first. That is not carelessness, it is how sampling works, and it is the same reason an image never regenerates identically.

  • Month three does not survive to month twentyA matter that runs two years is dozens of sessions. Each one starts from nothing, so the analysis you relied on in month three cannot be produced again in month twenty. You can keep the output. You cannot get back to how it was reached.
  • Your position appears to move, with no record of whyOpposing counsel asks why the argument changed between the demand letter and the brief. The honest answer is that the system was asked twice and answered differently. That is a difficult sentence to put in front of a judge.
  • The reasoning leaves when the associate doesThe person who ran the research in year one is gone by the time the matter is tried. What they knew about why one line of authority was dropped was never written down anywhere the firm can reach.
  • The firm pays twice for the same workSettled analysis gets re-derived because nothing carries forward. That cost is invisible on an invoice and it compounds on exactly the long matters that are worth the most.

Marcella keeps the matter, not just the answer. Every query, every citation and every analysis is written to the firm’s own Microsoft 365 tenant and stays attached to the matter it belongs to. Month twenty can read month three.

The problem

The matter is in four systems. None of them talk.

  • Document managementSharePoint, iManage, NetDocumentsThe drafts, the exhibits, and the version that was actually filed.
  • Research platformWestlaw, Lexis, vLexThe authorities, and whatever was checked before signing.
  • Practice managementClio, MyCase, PracticePantherThe deadlines, the parties, and the billing narrative.
  • Email and chatOutlook, TeamsThe decisions, and the reason behind each one.

None of the four can answer a question that spans the others. So a person assembles the matter by hand, for the partner, for the paralegal who picks it up next, and for the file.

That assembly is the work, and it is thrown away. The next person opens the same four systems and does it again, and the reasoning that got there the first time is not written down anywhere they can read it.

Pasting the matter into a chat window does not fix it, and neither does the same model on the firm's own API key. The second changes who holds the key. Neither one leaves anything behind that the next person can open.

A shared AI workspace moves the problem rather than removing it. It becomes a fifth place the matter lives, with its own copy and its own syncing back to the systems of record.

Marcella does not install a connector into any of them. It changes where the assembled context lives and how long it survives: in the tenant your firm already has, scoped to the matter, open to whoever picks up the file next.

01 / 04

Private to your tenant

From the Small band up, your matters live in the Microsoft 365 tenant your firm already has. Not on our servers. There is no cap on how many matters you keep.

02 / 04

Privilege by architecture

Client data stored on a third-party cloud may compromise privilege. Matter memory in your own tenant eliminates that risk. Architectural, not contractual.

03 / 04

Shared across the firm

Shared matter memory starts at Small, $750 a month flat for one to fifteen attorneys. One attorney's research reaches the next instead of being done twice. At Medium, $1,500 for sixteen to fifty, it is shared across offices and runs in your own Azure subscription.

04 / 04

Your 200th matter is stronger than your first

Marcella remembers which arguments prevailed in your jurisdiction, which citations your judges relied on, which opposing counsel patterns emerged across your caseload. That institutional knowledge compounds. It does not reset.

Read the answer before you trust the vendor.

The block above each answer is assembled from what retrieval found, not written by the model.

In the document

See which citations resolved, inside Word.

Two of three authorities were retrieved and carry a source anchor. The third does not exist in any reporter.

Back to top