MemoryBeta
Velt learns from past reviews. Previous approvals surface as precedent, so the tenth review is consistent with the first as teams grow.
No more review features a competitor can clone in a quarter.
Free tier. No credit card. First comment in 5 minutes.
I found a prior decision on this clause: indemnity cap was reviewed and approved in March. Not re-flagging.
precedent · clause 12 indemnity
Approved by Maya (Legal) · 14 Mar 2026 · cap set at 2× annual fees; no further liability.
source: contract-v1.pdf · thread #2281
context loaded · 3 items
Indemnity clause matches the Mar 14 approval: skipping. New pricing claim on slide 6 has no precedent.
Review intelligence running inside products at
What it is
Velt Memory is a learning layer over your users’ review history. Past decisions surface as precedent — to review agents before they review, and to human reviewers in context — so the tenth review is consistent with the first. The precedent accumulates in your product: a competitor can clone your features in a quarter; they cannot clone your customers’ settled judgments.
AI suggestion
Recommend approve · Confidence 91% · based on 47 similar judgments
every entry shows who decided, when, and why
both actor types feed the precedent; the human makes the call
How it works
Enable Memory and your review activity becomes structured judgments automatically. Search, ask, and suggest over them through the REST API; agents and humans read the same precedent.
npm install @veltdev/react<VeltProvider
apiKey={VELT_API_KEY}>
<YourApp />
</VeltProvider>// enable Memory in your workspace first
import { useSetDocument } from "@veltdev/react";
useSetDocument("filing-q3");Enable Memory in your workspace and Velt starts recording review activity as structured records. Each decision is enriched into a judgment automatically: the decision, the reasoning, who decided (human or agent), and the content type. Judgments are read-only through the API — there is no create endpoint — so a new workspace starts empty and fills as reviews happen. Search runs over two embedding spaces (the decision with its reasoning, or the reviewed content), filterable by decision, judge type, content type, reviewer, and date. Ask returns a grounded answer with citations and a confidence score, or an empty answer when nothing relevant exists. Suggest returns a recommendation with confidence and the judgment count it drew on. Memory recommends; it never acts. The reviewer always decides, and that decision becomes the next record.
// judgments are read-only · a new workspace starts empty and fills as reviews happen
Teams that build this budget a year of ML infrastructure work and keep paying for eval upkeep. The 3 steps above replace the year; once Memory is enabled, your review activity builds the record from the first decision.
Skip the steps. Have your agent set it up.
npx -y @velt-js/mcp-installer
claude mcp add velt-installer -- npx -y @velt-js/mcp-installer
npx -y @velt-js/mcp-installer
npx -y @velt-js/mcp-installer
npx -y @velt-js/mcp-installer
Knowledge sources
Powers these Velt surfaces
Launch review memory this week.
No credit card. Works with React, Next.js, Vue, Angular, and HTML.
Capabilities
Each card shows the real mechanics. Toggle to Code for the exact snippet behind it.
const ctx = useInlineContext({ documentId });
// → similar judgments, who, when, whyBefore a reviewer opens the filing, Memory surfaces similar past judgments: 3 approved last month, who decided, and why. The tenth review starts where the first nine ended.
POST /v2/memory/suggest
{ "documentId": "filing-q3" }Based on 200 judgments: recommend approve. Confidence 94%. Every suggestion carries its source records, so routine items clear in seconds and low-confidence ones get a full read.
POST /v2/memory/search
{ "query": "indemnity clause", "space": "decision" }Two embedding spaces index every judgment: the decision with its reasoning, or the reviewed content itself. The deck that repeats a settled mistake surfaces its history.
7 projections were rejected in Q3: each for variance over the declared 15% threshold.
POST /v2/memory/ask
{ "data": { "question": "rejections of Q3 projections" } }Ask in plain English: show me all rejections of financial projections in Q3. The answer cites the records behind it, so the audit takes minutes, not a quarter-end scramble.
POST /v2/memory/knowledge/ingest
{ "file": "brand-guide.pdf" }Upload the 50-page brand guide, the policy PDF, the pricing spreadsheet. Memory converts each into searchable text and extracted rules, so reviews check against the actual guide.
GET /v2/memory/knowledge/rules
?sourceId=qa-checklistIngest the 300-item QA checklist; Memory extracts it into versioned, citeable rules. Updates produce a diff, and rules nobody enforces anymore surface for cleanup.
POST /v2/memory/search
{ "query": claim, "judgeType": "agent" }Review agents query past judgments and knowledge before flagging. What the org approved before is not re-flagged; what it rejected is caught earlier. A human still decides every finding.
GET /v2/memory/alerts
?type=standards-driftMemory compares stated rules to actual behavior: the brand guide says sentence case, 60% of approvals use title case. The gap surfaces before the client or the auditor finds it.
projections > 15% variancealways human readvelt.memory.declare({
rule: "variance > 15% → human read",
});Declare it once: brand voice changed in January; projections above 15% variance always get a human read. Declared rules override inferred patterns, so policy beats habit.
GET /v2/memory/profiles/get
?reviewer=mayaEach reviewer’s history becomes a behavioral profile: approval rate, average review time, top flags raised. The contract goes to the reviewer who actually clears contracts.
Routine brand reviews used to take a day of back-and-forth. With precedent in context, the reviewer sees what we approved before and clears the obvious ones in seconds.
Head of content ops · enablement platform
Little big details
What ships today leads; pre-GA and spec-only capabilities render with a Coming soon tag, never in shipped tense.
Make it yours
Inline context panel, suggestion card, and NL query results as components for the fast path, with a full REST surface underneath for backend and agent integrations.
Inline context panel, AI suggestion card, and NL query results as components; panel position configurable (sidebar, bottom, modal); CSS variable theming, white-label, dark mode; headless hooks for fully custom UIs.
<VeltInlineContext position="sidebar" />
// or build on useMemorySuggestion()Full REST API for backend and agent integrations: search, ask, suggest, judgments query, knowledge lifecycle, profiles, patterns, stats, and alerts. Alert frequency, types, and severity thresholds via alert config.
POST /v2/memory/alerts/config/update
{ "maxAlertsPerWeek": 3 }We pointed our review agent at the suggest endpoint and it stopped re-flagging settled clauses overnight. Reviewers only see what is genuinely new.
Founding engineer · AI contract review
In production
Tabbed by vertical. Pre-GA, these are pilot patterns.
This Digital Sales Room gives you everything you need to move forward — from proposal to plan.
@Fin We need to make a working demo for Rene and his team
The client’s banned claims and settled brand debates surface as precedent on the next deck, email, and page. Brand review stays consistent when the content team doubles.For sales enablement
finance / receivables
Can we confirm the PO before approving this one? @Sarah
1 ReplyPast sign-offs on budgets, forecasts, and filings become queryable precedent with reasoning attached. The auditor’s question returns cited records in minutes.For fintech and FP&A
Cotton Supplier Ltd.
AgriCotton Farms
Village Kheda, Gujarat, India
SpinCo Yarns
18 Narol Rd, Gujarat, India
DyeWorks International
Calle de Mayo, Mexico
Can you attach the GOTS cert for this lot before we approve? @Sean
1 ReplyDecisions on orders, shipments, and field records accumulate into the org’s actual standard. The new coordinator reviews like a five-year veteran.For operations
Privacy Risk Review
Wren discovered a feature request in Jira involving users' health data and triggered a review after analyzing the PRD in Confluence against company policy and US regulation.
JStarted by Juan MendezThis feature processes PHI, so HIPAA applies. I recommend a PIA and a vendor BAA review before any data flows. @Juan
Past attestations and settled control decisions surface as precedent on the next filing. The examiner’s question returns cited records with the reasoning attached, in minutes.For compliance
Superflow
1 · Subscription & Fees
Customer subscribes to the Service on a monthly basis and shall pay the then-current fees in advance of each billing period. Fees are non-refundable except as expressly set out in this Agreement.
2 · Term & Auto-Renewal
The subscription renews automatically for successive one-month terms unless either party gives at least thirty (30) days' written notice of non-renewal before the end of the then-current term.
3 · Limitation of Liability
Except for breaches of confidentiality, each party's total aggregate liability under this Agreement shall not exceed the fees paid by Customer in the one (1) month preceding the event giving rise to the claim.
4 · Governing Law
This Agreement is governed by the laws of the State of Delaware, without regard to its conflict-of-laws provisions, and the parties consent to the exclusive jurisdiction of its courts.
Prior clause positions and the redlines counsel accepted last time surface on the new matter. The team stops re-litigating settled language, and the playbook holds as headcount grows.For legal
Make the hero headline punchier and add a primary CTA.
Done — rewrote the headline, tightened the subtext, and added a primary CTA in hero.tsx.
Add a logo strip with social proof under the CTA.
Added a five-logo trust strip below the CTA and wired it into hero.tsx.
Ship your app in a weekend
Get startedUse our brand orange on the CTA and cut the headline to one line. @Sean
Agents read the decision history before reviewing generated work, so settled decisions stay settled. Every suggestion carries confidence and provenance.For AI-native SaaS
See it running in products like yours.
30 minutes, with an engineer, not a sales deck.
Related primitives
Comments and approvals create the records; Memory turns them into precedent.
Memory is what makes them accurate: grounded agents check against the org’s actual standards.
Reads org standards before flagging, so findings land accurate.
The immutable records Memory indexes and learns from.
Routing gets smarter with Memory’s reviewer and content data.
Review intelligence is isolated per customer with physical namespace separation: a foreign sourceId returns empty results and never a cross-tenant read. GDPR deletion, legal hold, and per-content-type retention are part of the model. Beta access requires being added to the beta list. See self-hosting and governance.
Book DemoProof
Accumulated review history became the reason customers stay.
10x
New reviewers ramp in days, not months. The precedent reviews with them, so their tenth review matches our most senior person’s first.
minutes
The compliance audit used to be a quarter-end scramble. Now every rejection is a cited query and the answer comes back in minutes.
0 re-flags
Our agent stopped re-flagging what we already settled. Reviewers trust it because every suggestion shows its sources and confidence.
FAQ
Install @veltdev/react, wrap your app in VeltProvider with your API key, and your review activity is recorded as structured records. Enable Memory and mount the inline context panel where reviews happen; precedent and suggestions surface automatically.
No. Each customer’s records live in a physically separate namespace, and nothing crosses it: not records, not content, not names.
Velt indexes the decisions made in your product: every approval, rejection, and comment with its reasoning. When similar content comes up for review, the relevant past decisions appear in context, like “3 similar items approved last month,” each with who decided, when, and why.
Yes. Agents query the same API: past judgments, knowledge sources, and extracted rules. A grounded agent does not re-flag what your org already settled, and its suggestions carry confidence scores and cited sources. A human still decides.
Brand guides, policy documents, QA checklists, and data sheets: PDF, CSV, Excel (.xlsx), and plain text. Files up to 5 MB upload inline; larger files up to 30 MB go through a signed upload URL. Checklists and guides are converted into versioned, individually citeable rules.
No. Memory suggests; it never acts. When it lacks grounding it returns nothing rather than guessing: a question with no relevant history comes back with an empty answer, not an invented one. The reviewer accepts or overrides, and either way the decision is recorded.
Nothing, by design. Once you enable Memory in your workspace, judgments are created automatically from review activity, not imported or written by hand, so it starts empty and gets richer as reviews happen. Until then, questions return an empty answer rather than a guess, and suggestions stay sparse.
Beta means the feature is real and running in production with two caveats: API contracts may still change between releases, and access requires being added to the beta list. Ask through Book Demo or your existing Velt contact.
Velt is priced on usage, not seats: you pay for documents with review activity in a month, and there is a free tier for development and early production.
Free tier. No credit card. First comment in 5 minutes.
30 minutes, with an engineer, not a sales deck.
@Jordan Auto-renewal needs 60 days' notice to match our policy, not 30.
@Maya Liability cap is one month of fees — Legal wants 12 months before we sign.
@Maya Governing law should be New York to match the MSA.