You can ask plain-English questions across your own signed NDAs and contracts, and that beats keyword search whenever the answer lives in a clause, not a word. Keyword search matches characters. It has no idea what a non-compete tail, an evergreen renewal, or a governing-law clause is, so the questions you most need answered come back either empty or buried.
Picture the folder. Twenty signed PDFs. Mutual NDAs, a couple of consulting agreements, a vendor master services agreement, two employment offers with restrictive covenants stapled to the back. You need one answer fast: does anything in here stop your new client from being a competitor next quarter? Type non-compete into the search box and you get the three documents that use the phrase. The document that buries the restriction inside a section labeled Restrictive Covenants, never using the word compete, stays invisible. Ctrl-F told you there were three. You believed it. That is the failure mode. Not that search returns nothing, but that it returns a confident, incomplete number.
The real problem: signed PDFs plus a search box that does not read law
Full-text search treats every keyword independently. It ignores how words relate and what surrounds them. That is the documented limitation, not an opinion. It matches words, not meanings, so it misses synonyms, variations in phrasing, and any concept a contract expresses across a clause instead of in a single term.
Contracts are built to defeat that exact behavior. The same obligation shows up as covenant not to compete in one agreement, restrictive period in another, and post-termination restraint in a third. A keyword engine sees three unrelated strings. A clause-aware question sees one concept. Here is what most contract-search guides will not tell you: the danger is not the empty result, it is the confident one. Ctrl-F gives you a confident number. It does not tell you the number is wrong. When your tool cannot tell that those three phrases mean the same thing, you read all twenty PDFs by hand anyway, which is the work you were trying to skip.
Five questions keyword search cannot answer
These are not noun-swaps on the same query. Each one fails for a specific structural reason. Read them as the questions a single lawyer asks about their own files, then notice why a search box chokes on every one.
1. Does any NDA hide a non-compete or non-solicit tail that outlives the secrecy period?
This is the buried-tail problem, and it is the one that costs you. An NDA can carry a term of one year while the confidentiality obligation survives for three, and a non-compete or non-solicit restraint folded into the same document can run on a different clock entirely. The term of the agreement, the survival of confidentiality, and the restrictive-covenant tail are three independent timelines that often live in three different sections.
Keyword search fails here twice. First, the restraint may never use the word non-compete. Second, even when you find the clause, the date that governs it sits in a survival or termination section, so matching the phrase tells you nothing about when the obligation ends. You have to read the relationship between two clauses, and character matching cannot do that.
2. Which signed agreements auto-renew, and when does the opt-out window close?
An automatic renewal clause, also called an evergreen or self-renewal clause, extends a contract for another term unless a party gives notice of intent not to renew before a deadline. That deadline is the whole game. Notice periods commonly land at 30 to 90 days before expiration, but real contracts range from as few as 21 days to as long as two years, and many states now regulate how these clauses must be disclosed.
Miss the window and you are locked into another term. Keyword search will surface the word renewal. It will not compute the date math, will not tell you which renewal triggers next, and will not flag the contract whose notice period sits in a definitions section three pages from the renewal language. The question you want answered is a sort and a deadline, not a string match.
3. Find every contract governed by a given state law or arbitration in a given venue.
Governing law, jurisdiction, and venue decide which rules apply and where a dispute gets resolved. The governing-law clause sets which jurisdiction's law interprets the contract. An arbitration clause can route disputes away from court entirely. These usually track the drafting party's home state, so a folder signed with different counterparties can span five governing-law states before you notice.
A keyword search for Delaware finds the contracts that name Delaware. It does not find the one that says the State of incorporation of the Disclosing Party and defines that state earlier. It cannot separate the substantive governing law from the law of the arbitration agreement, which can differ inside the same document under the doctrine of separability. You want a grouped answer by jurisdiction, not a pile of hits.
4. Show my indemnification and limitation-of-liability clauses side by side.
These two clauses do opposite jobs, and they get conflated. A limitation-of-liability clause caps how much one party can recover. An indemnification clause shifts responsibility for specific losses, often third-party claims, from one party to another. They are complementary, and an indemnity can require payment beyond the liability cap, which is why you want to read them together rather than one at a time.
Keyword search cannot lay them side by side. Search indemnify and you miss the clause that says hold harmless. Search liability and you drown in every sentence that happens to use the word. Putting each contract's two clauses next to each other is a comparison task, and comparison is a question, not a query string.
5. Which contracts carry perpetual trade-secret obligations versus a fixed term?
Many modern NDAs split the clock on purpose. General confidential information gets a fixed term, commonly two to five years, while trade secrets stay protected for as long as they remain secret, because the Defend Trade Secrets Act and state law under the Uniform Trade Secrets Act protect information whose value comes from being secret. One document can hold both a three-year obligation and a perpetual one, side by side in different sentences.
Character matching cannot tell the two clocks apart. The word confidential shows up in both clauses. The word perpetual may not appear at all, because the perpetual obligation is usually phrased as for so long as the information remains a trade secret. Sorting your folder into fixed-term versus indefinite is a reading-comprehension question, and that is the line keyword tools cannot cross.
Why keyword search fails each one, in one view
| The question | Why keyword search fails | What you actually need |
|---|---|---|
| Buried non-compete or non-solicit tail | Restraint avoids the word compete and its expiry sits in a separate survival section | Read the link between the covenant and its own timeline |
| Auto-renewal and opt-out deadline | Matches the word renewal but cannot compute the notice window or rank what triggers next | A sorted answer with each opt-out deadline |
| Governing law or arbitration venue | Misses defined-term references and cannot split substantive law from arbitration law | Contracts grouped by jurisdiction |
| Indemnity vs limitation of liability | Misses synonyms like hold harmless and floods on the word liability | The two clauses placed side by side per contract |
| Perpetual vs fixed-term secrecy | The word confidential appears in both clocks; perpetual is phrased indirectly | Each obligation labeled fixed term or indefinite |
This is a real research task, not a wish
Asking a yes/no question about an NDA clause has a formal name in computer science: document-level natural language inference. Stanford researchers Yuta Koreeda and Christopher Manning built a dataset for exactly this in 2021, called ContractNLI. They took 607 real NDAs and wrote 17 fixed hypotheses, things like the receiving party may share confidential information with third parties, then trained systems to label each one as entailed, contradicting, or not mentioned, and to point at the evidence span in the contract. That is the same shape as the five questions above. The legal questions you would ask a colleague map directly onto a task machines are already measured on.
Ask, do not search
The shift sounds small. Stop typing terms, start asking questions. Natural-language retrieval over your own corpus lets you write the question the way you would say it to a colleague. Which of my NDAs has a non-solicit tail that runs longer than the confidentiality term. Which agreements renew automatically in the next ninety days. Group my signed contracts by governing law. The system reads the clauses and answers from the documents, instead of forcing you to guess the magic word the drafter happened to use.
This is retrieval, not judgment. A question-driven tool can find the clause, surface the date, and put two provisions next to each other. It does not tell you whether the restraint is enforceable in your state or whether to sign. Keep the claim where it belongs: faster finding, faster reading, the same lawyer making the call.
Retrieval finds and arranges the clauses. It does not give legal advice. Whether a non-compete tail is enforceable, or a renewal still binds you, is a judgment for a licensed professional, not a search result.
Personal, not a workspace
This is deliberately not contract lifecycle management. CLM platforms are built for teams, with repository onboarding, approval routing, and shared workspaces, and they assume an organization feeding them documents. The persona here is one person and their own files: a solo practitioner, an in-house lawyer keeping a personal copy of what they signed, a founder with a drawer of executed PDFs.
No onboarding project. No firm-wide index. No other people's data. You drop your own signed agreements into a personal space and ask questions against them. The scope is your corpus, which is what makes the ask-do-not-search model practical for an individual instead of an enterprise rollout.
Where MemX fits
MemX (memx.app) is a consumer AI memory app, a personal second brain for your own saved material. You capture PDFs, documents, photos, voice notes, and WhatsApp messages, then retrieve by asking in plain English. MemX answers from your saved material and cites the source memory, so when you ask which of your NDAs mentions a non-solicit period, it points you back to the exact document. Its OCR reads text inside scans and photos, so a signed agreement you photographed is searchable by its contents, not just its filename.
MemX is private by architecture: customer-managed encryption keys, per-user cryptographic isolation, encryption at rest, and on-device processing options. It is an app that helps one person search their own saved documents by asking. It is not enterprise eDiscovery and it does not reach into anyone else's files. There is a free tier with no credit card on Android, iOS on the App Store, and WhatsApp. For the lawyer with a folder of their own signed PDFs and a question, that is the right size of tool.
When you capture a signed PDF, ask one verification question right away, such as what is the governing law of this agreement. If the answer cites the right clause, you know it indexed cleanly before you need it under deadline.
FAQ
01Can I search across my contracts and NDAs by asking a question instead of keywords?
Yes. Natural-language retrieval reads clauses rather than matching strings, so you can ask which agreements auto-renew or which carry a non-compete tail and get an answer drawn from your own documents, with the source clause cited.
02Why does keyword search miss a non-compete buried in an NDA?
Full-text search treats words independently and matches exact terms. A restraint phrased as a restrictive covenant, never using compete, will not surface, and its expiry often sits in a separate survival section the search never connects.
03How do I find every contract governed by a specific state's law?
Ask the question directly and have the tool group results by jurisdiction. Keyword search misses defined-term references and cannot separate the substantive governing law from the law governing an arbitration clause inside the same document.
04How long does an NDA confidentiality obligation last?
General confidential information commonly gets a fixed term of two to five years. Trade secrets stay protected for as long as they remain secret under the DTSA and state UTSA law, so one NDA can carry both a fixed clock and a perpetual one.
05What is the difference between indemnification and limitation of liability?
Limitation of liability caps how much one party can recover. Indemnification shifts responsibility for specific losses, often third-party claims, and can require payment beyond the cap. They do opposite jobs and are best read side by side.
Written by Aditya Kumar Jha, who works on AI memory at MemX, where the product is simple: your own documents, answered by a question. None of the above is legal advice. For whether a clause binds you, ask a licensed lawyer; for where the clause is, ask your files.
