--- title: "Memory for legal and compliance AI assistants" description: "Memory for legal and compliance AI: how an assistant answers which version of a clause or policy governs today, cites the document behind it, and reports when the record does not support an answer." canonical: https://past.dev/use-cases/legal last-updated: 2026-09-01 --- # Memory for legal and compliance Source: https://past.dev/use-cases/legal A legal or compliance assistant answers questions where the current version is the whole answer: which clause governs after three amendments, what the retention policy required in May, whether this counterparty is under the old terms. Memory for legal and compliance AI has to order versions by date, cite the document behind each answer, and decline when the record does not support one. This page describes what that takes. ## What you are building An assistant over a firm's own documents and the external material it works against. Internally that is contracts and their amendments, policies and their revisions, regulatory correspondence and matter files. Externally it is regulations, regulator guidance, standards, industry codes and counterparty paper. Anything you can send as text with a date joins the same record, so a question can be answered across both at once. It answers questions from lawyers, compliance officers or employees, and every answer it gives is one somebody may rely on. > **Scope** > > past.dev stores evidence and returns it with dates and sources. It does not perform a legal or compliance judgement, and nothing here is a substitute for one. The judgement stays with the firm. ## Why memory decides whether it works - **Three amendments, one clause, no ordering.** The original and every amendment are all retrievable and all plausible. Similarity ranking has nothing that says which one governs today. - **A superseded policy answers as current.** A retention rule revised in June sits in the record beside the version it replaced. Quoting the old one is a finding at the next audit. - **One counterparty, three names.** The MSA, the DPA and the email thread name the same entity three ways, so a question about the relationship answers from a third of it. - **An uncited answer is unusable.** In a regulated review, a claim without the document behind it cannot be relied on, filed or defended. - **A confident guess is the worst outcome.** An assistant that composes a plausible clause when the record holds none has manufactured a term. ## What past.dev provides - **Version ordering by date.** The clause in force is the one supported by the most recent evidence, and every superseded version keeps its own date, so what governed on a given day stays answerable. See [fact supersession](/glossary/fact-supersession) and [validity windows](/glossary/validity-window). - **Citations on every answer**, back to the stored document and the sentence within it, with the date it was written. - **An explicit insufficient-evidence status.** When the record does not support an answer, the response says so rather than composing one. For this use case that property does more work than any other. - **Matter separation through audiences.** A data point carries an `audience` list and retrieval filters on it inside the engine, so a shared assistant does not reach a matter it should not. Audiences are configured with past.dev during early access. - **Deletion that reaches derived facts**, so removing a source removes what was extracted from it, which is what an erasure request requires. - **A record of every call.** Each request to the developer API is written to the request log with its outcome, including refusals, and the console shows them by project. ## Cost and accuracy A matter with ten years of documents costs about the same per question as one opened this week, because recall is not metered and query cost does not scale linearly with stored history. The alternative, pushing a document set into the context window on every question, is priced per question and gets more expensive exactly as a matter gets older and more valuable. Accuracy here is not a preference. An assistant that quotes a superseded clause, or invents one, is worse than no assistant, because somebody acts on it. Dated supersession and an explicit insufficient-evidence status are the two properties that address it, and both are in the response rather than in prompt instructions. Measured results are on [benchmarks](/benchmarks) with the method at [how to reproduce it](/benchmarks/methodology). Certifications and the data processing agreement are on [security](/security). ## Getting there 1. Send internal and external material to `POST /api/v1/ingest` through the same call: contracts, amendments and policies alongside the regulations, guidance and standards they are read against. Give each the date it was executed, issued or published rather than the date you loaded it. 2. Set an `audience` per matter or per practice area at ingestion, so retrieval separation is enforced in the engine. 3. Call `POST /api/v1/recall` and pass only the returned evidence to the model, so the answer is bounded by what the record supports. 4. Route on the status: answer with the citation and its date when supported, surface both versions when the evidence conflicts, and report insufficient evidence otherwise. ## Frequently asked questions ### Can it answer what a policy required on a past date? Yes. Every observation is stored with the date it happened and a superseded version keeps its date rather than being overwritten, so the state on a given day is a question the record can answer. ### What stops it inventing a clause that is not in the documents? Recall returns ranked evidence and a status. When nothing in the record supports the question the status says so, and the application can refuse rather than answer. The application decides what to do with that status, so refusal has to be wired in deliberately. ### Can a shared assistant reach a matter it should not? A data point carries an audience list and retrieval filters on it during the search, so restricted material is not loaded into an answer. Audiences are configured with past.dev during early access rather than self-serve. ### Is past.dev giving legal advice? No. It stores documents and returns dated evidence with citations. Every legal and compliance judgement stays with the firm, and the assistant you build on top is yours. ## Related - [Knowledge base](https://past.dev/use-cases/knowledge-base) - [Customer support](https://past.dev/use-cases/customer-support) - [Security](https://past.dev/security) - [Benchmarks](https://past.dev/benchmarks) - [Quickstart](https://past.dev/docs/memory-api/quickstart)