The Case for Giving Legal AI a Memory: What SimpleDocs Shipped in April

The release I keep coming back to when I reflect on what we shipped in April is a small one called AI Memory. It lets the assistant remember the durable preferences you tell it and apply them automatically afterwards, opt-in per user and off by default. On its own it is a convenience. Taken seriously it points at the thing legal AI has been missing all along. A model that forgets everything between sessions cannot hold your standards, your precedent, or your way of working. Memory is the precondition for context, and context is the precondition for output a lawyer can actually rely on.

Standards only work when the right people own them

Playbook Permissions shipped this month: you can now share a playbook with specific people or groups and control who can view or edit it. That sounds administrative, and it is, but it is also the difference between a playbook being one person's document and being the team's standard. A standard with no owners drifts. A standard anyone can quietly edit is not a standard either. Permissions are how a written position stays the position.

Further, a major concern is of senior counsel leaving a team, and taking their playbooks with them.

Alongside opt-in memory, Custom AI Rules let an individual save their own instructions and have them applied automatically across chat and review, and playbook review itself became more accurate, with fewer false positives when it evaluates an instruction. The Improve Instructions tool now works on fallback positions as well as the primary one, which is where most real negotiation actually lives.

The unit of comparison should be the whole agreement

Precedent Benchmark, Full Document lets you benchmark an entire agreement against precedents rather than checking one clause at a time, choosing which clauses to include and adding others from the document for comparison.

Most of the repository work was unglamorous, and that is the point

A lot of April went into the foundational parts of repository nobody demos.

• Saved Views let people save, reopen, and share custom document views.
• Repository Insights added metrics and trend reporting. Imports improved in several quiet ways: documents from SharePoint can carry their library columns in as repository fields, there is now a history page for smart imports, and admins can bulk update field settings through a spreadsheet.
• Extraction became more accurate on the fields people actually filter on, with stricter handling of boolean values and better source attribution back to the document.

None of that is exciting to write about, but it's what determines if the precedent layer can be trusted. A repository full of slightly wrong values is worse than no repository.

Memory is off by default

We would rather it earn trust before it shapes anyone's work without their noticing. The same instinct runs through the month: the model was never the hard part. Deciding what the model should remember, and who gets to set it, is the work that actually belongs to a legal team. We spent April building the places that decision gets recorded.

I have silently begun work on oneNDA_v3.0 this month too, updated for the AI data usage era, which has seen us poring over 10,000+ AI policies and documents to identify the standards which are beginning to form. I am excited to share more of this in due course.

Also shipped this month

• Signature CC: auto-copy a configured address on signature request emails.
• Saved View filters: adjust, disable, override, reset, and persist filters.
• AI Templates: a simpler create, tag, save, and fill workflow.
• Context-aware redlines: AI review now draws on defined terms and referenced clauses.
• Faster clause review: standardised extraction and auto-loading batches.
• Per-company AI provider: route a company's AI work to a chosen provider (in a chosen region).
• Faster access for users in China (we operate in 89 countries).
• Paid Law Insider legacy accounts gained access to the SimpleAI Word add-in.