Revo Memory

Memory built for organizations.

Your company's context through an API. Every data point ingested, encrypted and permissioned, then served to humans and agents on demand.

Request early access

Read the docs

on BEAM, at every context size

23x

cheaper per question than sending 1M tokens of history

Quickstart

Three calls. No schema, no SDK, no index to build.

One input shape: text plus a timestamp. You do not pick an embedding model, chunk anything, maintain entity tables or write prompts that decide which of two conflicting statements is current. Those are memory operations, and they happen on our side.

# 1. Push a fact. Timestamps are backdatable, so history lands on its real dates.
curl -X POST https://api.revo.ai/ingest \
  -H "Authorization: Bearer $REVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "content": "Nadia confirmed the pilot ships March 14. Budget is 32k.",
        "label": "Account review call",
        "timestamp": "2026-02-03T09:00:00Z",
        "group": "acme-pilot" }'

# 2. Push its update, five months later.
curl -X POST https://api.revo.ai/ingest \
  -H "Authorization: Bearer $REVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "content": "Budget for the Acme pilot moved to 40k.",
        "label": "Email from Nadia",
        "timestamp": "2026-07-28T16:00:00Z",
        "group": "acme-pilot" }'

# 3. Ask. Returns 40k, with the July email as evidence and the February
#    value kept as dated history.
curl -X POST https://api.revo.ai/answer \
  -H "Authorization: Bearer $REVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "query": "what is the Acme pilot budget?" }'

Full quickstart API reference

Benchmark

The strongest recall in AI memory.

Every vendor publishes a score it graded itself, so those numbers cannot be compared. We ran ours through their setups instead: Exabase's published methodology, and Hindsight's harness taken verbatim with only the API calls swapped. We lead on both.

76.9%
Exabase M-1
2nd
87.3%
Revo Memory
1st
73.4%
Hindsight
3rd

Recall accuracy on BEAM at 100k tokens. Our score is the mean of two runs on our competitors' own setups.^*

*Numbers published by different vendors are not comparable: each grades its own answers, with its own judge, on its own harness. Ours is the mean of two runs on their setups, Hindsight's harness taken verbatim and Exabase's published methodology, because we lead on both and picking the higher one would be the behaviour this page argues against. Our own judge is stricter than either and puts us at 74.1% on the 100k split; we report that too.

What makes the difference

Privacy, entity resolution, temporality.

Retrieval quality is where the benchmarks look. These three are where a memory layer either works inside a real business or does not.

Privacy

Permissions applied while retrieval runs

Every piece of content carries an audience, and the filter applies during the search rather than to the results afterwards. Content the asker may not see is never loaded, so it cannot leak through a summary or a citation.

The workspace is resolved from your key on the server rather than passed as a parameter, so nothing in a prompt can redirect a call to another tenant. Deleting a source removes the facts derived from it and the sentences that produced them.

Entity resolution

One customer, however many ways they appear

An email address, a name in a transcript, a Slack member ID and a new address after someone changes jobs all resolve onto one node as evidence accumulates.

This is what makes memory work for an organization rather than one person. The same customer met by three of your users across email, a call and a ticket is one relationship with three sources, and questions about them return all three.

Temporality

Facts carry dates, and old values stay

A budget agreed in February and revised in July has one current value and a history. Ask what it is now and you get July, with its source. Ask what it was and February is still there.

Where sources genuinely disagree, both stay visible and both stay dated. A vector store returns the February and July lines side by side and leaves your model to guess which one still holds.

How it works

From raw text to a sourced answer, in five stages.

Every piece of text goes through the same pipeline, whether it is an email, a transcript or a PDF. POST /ingest runs the first three, POST /recall stops at stage four, and POST /answer goes on to stage five.

1
Capture
The original text is kept word for word, with its timestamp.
2
Understand
People, companies and projects are each resolved to one node.
3
Distill
A new value replaces the old one; the earlier value stays in history.
4
Recall
A question pulls everything relevant, ranked and permission filtered.
5
Answer
Grounded in the retrieved evidence only, returned with that evidence.

Pricing

You pay for what you ingest and what you ask.

The people using your software never need a Revo account, so we count what the API does rather than how many people touch it.

Usage

Metered

published rates, billed monthly

You pay for what you ingest, what you ask, and what the graph maintains. Every call returns its own cost under usage, so you can price a feature or cap a customer without instrumenting anything.

Metered on ingestion, questions and resolved entities
Included allowance, then per unit overage
All four endpoints
Answers with evidence and per call cost
Multiple workspaces

Enterprise

Custom

annual agreement

For teams with a procurement process. Everything in Usage, plus the operational and contractual terms a platform team asks for before they put your memory in front of their own customers.

Dedicated or in region deployment
SLA and support terms
DPA, retention policy and legal hold
Audit logging
Custom ontology for your domain

What counts as a unit

Security

Your data stays yours.

Your data is stored to answer your questions and is never used to train models. Deleting a source removes the facts derived from it along with the sentences that produced them. The platform is SOC 2 Type II certified, with ISO 27001 and ISO 27701, and the API runs inside that same scope.

Security overview

FAQ

The questions platform teams actually ask, including the ones where the answer is not yet.

How is this different from mem0 or Zep?

Both scope memory to a user. Revo scopes it to an organization: the same person, company or project accumulates evidence across every employee and every source, and resolves to one node rather than one per observer. Zep is the closest architecturally, since it also builds a temporal graph and invalidates superseded facts.

What do I have to build before I can send anything?

Nothing. There is one input shape, text plus a timestamp. No schema to define, no embedding model to pick, no chunking pipeline, no entity tables and no alias reconciliation. Four endpoints and one bearer token.

What happens when a fact changes?

The new value becomes current and the old one stays in the history with its date. Ask what is true now and you get the current value with its source; ask what it used to be and the earlier value is still there. Where sources genuinely disagree, both stay visible and both are dated.

Can I scope memory to my own end users?

Not yet through the API. The engine filters retrieval by audience already, but POST /ingest does not accept an audience parameter, so everything pushed in through the API lands with the workspace as its audience. That parameter is the next milestone on the roadmap. Until it ships, do not put data with different authorization boundaries into the same workspace.

Is there an SDK?

Not yet. There is no Python, TypeScript or C# client today, so you call the endpoints directly. There is also nothing to host and no schema to design, which is why the whole integration is a handful of HTTP calls.

How do I get a key?

By talking to us. Tokens are issued by hand during early access; there is no dashboard, no programmatic issuance and no rotation endpoint yet. Self serve keys are dated on the roadmap.

How do you measure retrieval quality?

On BEAM, a public benchmark from ICLR 2026 whose paper, dataset and harness are all public. We report two numbers per ability: whether the answer was inside what we retrieved, and whether a model then got it right from that evidence. We publish all ten abilities including the one we score worst on.

Is my data used to train models?

No. It is stored to answer your questions and nothing else. Deleting a source also removes the facts derived from it and the sentences that produced them.

Tell us what you are building.

Keys are issued by hand while the API is in early access. Bring your volumes and the shape of your integration, and we will tell you straight whether it fits yet.

Request early access

Read the docs