Aiolum Browse stories
All stories
Memory layer for AI agents in a single file · United States

Memvid: Revenue, team and how it was built

Small teamFundedB2BDeveloper tools

Memvid (Memory layer for AI agents in a single file, United States): US$20K a year in recurring revenue, July 2026. Evidence grade D: reported, not verified. Team: 2 co-founders; total headcount not published, November 2026. Launched 4 January 2026; still running.Added 7 November 2026

Outcome
US$20K a year in recurring revenue, July 2026
Evidence
D reported, not verified
Launch
4 January 2026
Status
Still running
Team
2 co-founders; total headcount not published, November 2026
Outside funding
Investment from Market Square Ventures and Brickyard in fall 2025, amounts not published; a US$800K pre-seed round listed as open by THIRD in July 2026
Main channel
Open source
Reported, not verified

What is not verified

The headline, US$20,000 in annual recurring revenue, rests on one source: THIRD's dealflow newsletter of 31 July 2026, a newsletter that offers investors introductions to startups that are raising money. It lists the figure with the company's other traction claims and does not say how it was checked. No payment-processor profile, founder post or press report with a revenue figure was found; the Teknovation interview of June 2026 gives no revenue. The user counts also differ: about 60,000 developers in June 2026 (the CEO to Teknovation) and more than 40,000 developer users in July 2026 (THIRD). A reader could look for a founder update with a dated revenue figure or a report on the US$800,000 round. The headline number is grade D: one claim or one estimate, with no second source. Aiolum shows it because the story is useful, and marks it so that it is not mistaken for a verified figure.

Memvid is memory infrastructure for AI agents: software that stores what an agent has read and done, so that the agent can find it again in a later session. Most such systems use a vector database, a server that stores embeddings (lists of numbers that represent the meaning of a text) and returns the texts whose numbers are closest to a question. Memvid instead writes the texts, their embeddings and the search indexes into one portable file with the extension .mv2, which an agent can open, search and append to without a server. Its search combines BM25, a classic keyword-ranking formula, with embedding search.

The company says it began when its two co-founders, Mohamed Mohamed (CEO) and Saleban Olow (CTO), built an AI agent for a healthcare staffing problem and found that its memory was unreliable. In 2025 Olow published a weekend prototype that stored embeddings in video frames; the founders wrote that it reached more than 10 million views and 10,000 GitHub stars. In fall 2025 the founders took investment from Market Square Ventures and Brickyard and moved to Chattanooga, Tennessee, Teknovation reported in June 2026. On 4 January 2026 they released Memvid V2, a rewrite in Rust with its own binary file format.

THIRD's dealflow newsletter, which lists startups that are raising money, reported on 31 July 2026 that Memvid had US$20,000 in annual recurring revenue, more than 40,000 developer users and paying customers, and was raising a US$800,000 pre-seed round. That is the only revenue figure found, so this story is grade D. The documentation lists paid plans at US$9.99 and US$49.99 a month in November 2026.

Memvid: US$20K a year in recurring revenue, July 2026. Memory layer for AI agents in a single file, United States, team of 2 to 10, evidence grade D.
The story in one card. Share it with the address of this page.
01

How much revenue does Memvid make?

Each row shows the evidence grade and the source numbers from the list at the end of this page. The grades are explained in How Aiolum verifies.

An open-source memory file for AI agents with paid storage plans; one reported revenue figure of US$20,000 a year in July 2026
FactValueGradeHow we knowSources
Annual recurring revenue, July 2026US$20,000DTHIRD dealflow newsletter, 31 July 2026, in a traction list next to "Paying customers today". The newsletter does not say how the figure was checked.[1]
Monthly equivalent of the July 2026 figureabout US$1,667calcCalculated: US$20,000 divided by 12.[1]
Revenue target for the end of 2026US$1 million a yearDTHIRD newsletter: "a target of $1M ARR by year-end". A target, not revenue.[1]
Round being raised, July 2026US$800,000 pre-seedDTHIRD newsletter. No source says whether the round closed.[1]
Earlier investmentMarket Square Ventures (Knoxville) and Brickyard (Chattanooga), fall 2025CTeknovation, 28 June 2026. Amounts not stated. The founders moved to Chattanooga to work at Brickyard.[2]
Developer usersmore than 40,000 (July 2026); about 60,000 developers had used the code within six months (June 2026)DThe first figure is from the THIRD newsletter; the second is the CEO's statement to Teknovation. The two counts do not agree and may measure different things.[1, 2]
GitHub repository memvid/memvid, 7 November 202617.4k stars, 1.5k forks, 261 commitsBRead on the public repository page. The THIRD newsletter gave 12,000+ stars in July 2026.[6, 1]
Reach of the first prototypemore than 10 million views and 10,000 GitHub starsCThe company's About page and Saleban Olow's Show HN post of 5 January 2026.[3, 7]
Founding team2 co-founders: Mohamed Mohamed (CEO) and Saleban Olow (CTO)CAbout page and THIRD newsletter. Teknovation writes the CEO's name as Mohamed Omar. The careers page speaks of "a small, focused team" and gives no headcount.[3, 1, 2, 11]
BaseChattanooga, Tennessee; the founders met in MinnesotaCTeknovation, 28 June 2026. The site footer names Memvid, Inc.[2]
Paid plans, November 2026Starter US$9.99 a month, Pro US$49.99 a month, Enterprise on requestCDocumentation page Capacity & Plans. The FAQ says paid plans go up to 100 GB per file; the plan table says 25 GB per memory for Pro. The two pages differ.[4, 5]
Open-source partcore engine (memvid-core) and command-line tool, Apache 2.0BLicense shown on the GitHub repository; the FAQ says the cloud dashboard and enterprise features are commercial. The CEO told Teknovation that the ingestion and retrieval pipelines are closed source.[6, 5, 2]
Release of Memvid V24 January 2026CBlog post by Saleban Olow: a rewrite in Rust with a custom binary format, QR codes removed.[8]
Default embedding modelBGE-small (384 dimensions), run locally; OpenAI embeddings optionalBREADME on the GitHub repository.[6]
Time from the V2 release to the July 2026 figure6 monthscalcCalculated: January 2026 to July 2026.[8, 1]

How to read this table: a grade D value comes from one source with one date, here a fundraising newsletter, and nobody has checked it. A grade C value is a statement by the company or its founders, in their own pages or in a news interview. A grade B value is visible in public code on GitHub. A "calc" value was computed on this page.

Memvid prices in November 2026, from its documentation page Capacity & Plans (grade C); the in-app plan page is closed to crawlers by robots.txt
PlanPer month, billed yearly (US$)Per month, billed monthly (US$)Limit
Free0050 MB, 1 memory file
Starter9.9925 GB total, 5 memory files of up to 5 GB
Pro49.99125 GB total, 10 memory files of up to 25 GB
Enterpriseunlimited; contact the company

The documentation gives monthly prices only and no yearly price. Plans differ by total storage, by the number of memory files and by the size limit of each file. The core engine and the command-line tool are open source under the Apache 2.0 license and can be self-hosted for free, according to the FAQ.

02

When did Memvid launch, and how did it grow?

From a weekend prototype on GitHub to a funded company with paid plans in about one year. Each entry carries its grade.

  1. 2024CMohamed Mohamed and Saleban Olow, who met in Minnesota, start working together on AI agents for healthcare systems. [2]
  2. Mid-2025COlow publishes a weekend experiment that stores text embeddings inside video frames. The founders say it reached more than 10 million views and 10,000 GitHub stars; the CEO said about 10,000 developers downloaded the code in the first week. [7, 3, 2]
  3. Fall 2025CInvestment from Market Square Ventures and Brickyard. The founders move to Chattanooga to work at Brickyard. [2]
  4. 4 January 2026CMemvid V2: a rewrite in Rust that stores memory in one .mv2 file with an embedded write-ahead log, a full-text index and a time index. [8]
  5. 5 January 2026COlow posts Memvid V2 on Hacker News as "SQLite for AI memory". [7]
  6. 28 June 2026CTeknovation reports that about 60,000 developers had used the code within six months and that the company plans a fundraising round in the fall. [2]
  7. 31 July 2026DTHIRD's dealflow newsletter reports US$20,000 in annual recurring revenue, more than 40,000 developer users and a US$800,000 pre-seed round being raised. [1]
  8. November 2026BThe documentation lists paid plans from US$9.99 a month; the GitHub repository shows 17.4k stars. [4, 6]
03

Why did Memvid work?

Four causes, in the order they mattered. Each one is a reusable lesson.

01

A free repository as the funnel

The first version was a free experiment on GitHub, and the founders say its reach brought thousands of developers before there was anything to buy. The paid plans sell storage on top of the same open-source engine, so each developer who tries the free code is a possible customer.

02

One file instead of a server

A vector database is a server that someone must run, pay for and back up. Memvid puts the data, embeddings, indexes and a write-ahead log (a record of changes that lets the file recover after a crash) into one file that can sit on a laptop, a private server or a cloud bucket. The company sells this as lower infrastructure cost and as privacy for data that must stay on site.

03

Open core

The engine is open source under Apache 2.0, while the CEO told Teknovation that the ingestion and retrieval pipelines are closed source. The cloud dashboard, larger capacity and enterprise features are what the company charges for.

04

Attention-seeking marketing

Teknovation reported that the company posted a job titled "AI Bully", for a person paid to chat with chatbots and record how they forget. The CEO said the stunt was picked up by large media outlets.

How it was built

These details come from the public GitHub repository, the company's documentation and blog, and the founders' statements. They are grade B for the code and grade C for the rest. [6, 7, 8, 5]

  • The engine of V2 is written in Rust, with SDKs for Python, Node.js and Rust, a command-line tool installed from npm, and an MCP server so that models such as Claude can read a memory file directly.
  • Each item is stored as a "Smart Frame" with its content, its embedding, a timestamp and its relationships, according to the Show HN post.
  • Search combines BM25 keyword ranking with embedding search. The company says search takes under 5 milliseconds on consumer hardware; this is its own benchmark.
  • The default embedding model is BGE-small, an open model from BAAI downloaded from Hugging Face and run locally in ONNX format; OpenAI's embedding API is an option.
  • The file contains a write-ahead log for crash recovery and a table of contents with a SHA-256 checksum for each segment, according to the V2 blog post.
  • Version 1 encoded text as QR codes in video frames. The README says it is deprecated and that a description of Memvid with QR codes is outdated.

Tools named in the sources

Each one opens its page in the Tools tab, with every story that uses it.

04

Who tried this before Memvid, and why did they fail?

Four companies in our database of failed startups tried to sell a new database, a search layer or a company memory to developers and businesses. The values below are from that database and are not independently verified.

Died 2009 to 2016Burned US$12M

RethinkDB

The promise

An open-source database that pushed query results to applications in real time.

Why it died

Developers liked it but would not replace Postgres for a feature they could build themselves.

No Market Need
Died 2020 to 2024Burned US$3M

WunderGraph

The promise

An open-source framework that joins GraphQL, REST and databases into one typed API.

Why it died

Open-source use did not turn into enough revenue; enterprise sales were slow while large infrastructure companies added the same features.

Ran Out of Cash
Died 2015 to 2021Burned US$4.5M

Omnity

The promise

Semantic search that connected documents across large, unstructured data sets.

Why it died

The cost of collecting and processing data was higher than the revenue, and large technology companies offered AI search themselves.

Competition
Died 2017 to 2020Burned US$2.8M

Niles

The promise

A Slack bot that stored a company's knowledge and answered staff questions.

Why it died

Slack added its own AI features, and cheaper competitors followed.

Competition

The failed startups come from our database of 1,749 dead startups, retold here in our own words. Its entries are AI-assisted summaries of public sources, so these cards carry no evidence grade.

What changed

RethinkDB and WunderGraph had open-source code that developers liked, and both struggled to turn that use into revenue. Omnity and Niles sold search and company memory and lost to larger companies that added the same features. Memvid faces both risks: its core is free to self-host, and model providers are adding memory to their own products.

What differs is the moment and the form. AI agents that run for days or weeks need memory, which few developers had to build before 2023. Memvid ships that memory as a file rather than a server, so a developer can start for free on a laptop, and the company charges for capacity and enterprise deployment. Whether that converts free users to paid ones is not shown yet: THIRD's newsletter reported US$20,000 a year in revenue in July 2026 against more than 40,000 developer users.

05

Can one person copy Memvid now?

The viral start, rarely. The open-core model, yes.

A weekend repository that reaches 10 million views cannot be planned, and the founders say they did not expect it. The structure can be copied by one developer: publish a small, useful library for AI agents for free, then sell hosting, capacity or a company version to the people who adopt it.

Where it is still open: Other parts of the agent stack that still need a server today: tool permissions, audit logs, evaluation data and offline search for private documents.

  1. 1Solve one problem that AI agents have in practice, such as memory, and publish the code under an open license.
  2. 2Explain it in one comparison that developers already know (here, "SQLite for AI memory").
  3. 3Keep the paid part separate: storage, a dashboard, access control, support.
  4. 4Publish a simple price table early, so that free users can see what an upgrade costs.

What could go wrong for Memvid?

The risks are the mirror image of the causes above.

What is not verified

The headline, US$20,000 in annual recurring revenue, rests on one source: THIRD's dealflow newsletter of 31 July 2026, a newsletter that offers investors introductions to startups that are raising money. It lists the figure with the company's other traction claims and does not say how it was checked. No payment-processor profile, founder post or press report with a revenue figure was found; the Teknovation interview of June 2026 gives no revenue. The user counts also differ: about 60,000 developers in June 2026 (the CEO to Teknovation) and more than 40,000 developer users in July 2026 (THIRD). A reader could look for a founder update with a dated revenue figure or a report on the US$800,000 round.

Logo wall without sources

The home page shows the heading "Trusted by builders at these companies" with logos of ByteDance, AWS, Google, Palo Alto Networks and Snapchat. No downloaded source says that these companies are customers; the words say only that builders at them use Memvid. The testimonials on the same page name their authors and link to their posts on X.

Competition

On this site, Moss sells a similar retrieval layer for AI agents, and Chatbase and DocsBot AI sell finished chatbots built on retrieval. Memvid's own home page compares it with Pinecone, Chroma, Weaviate and Qdrant, and THIRD's newsletter names Mem0 and OpenAI's memory as the systems it benchmarks against. Its speed and accuracy figures are its own benchmarks.

Dependence on model providers

Low for the core: the README says the default embedding model is BGE-small, an open model that runs locally, with OpenAI embeddings as an option, and the Show HN post says the file works with GPT, Claude, Gemini and Llama. A risk of a different kind is that model providers build memory into their own products, which the newsletter's comparison with OpenAI's memory shows.

Free self-hosting and an enterprise service

The FAQ says the core engine can be self-hosted for free forever, which limits what the paid plans can charge for. The enterprise product page promises a system "deployed end-to-end by our expert team" in one to two weeks, so enterprise sales include service work, although the product sold to developers is software.

Origin and naming

The About page says the founders first built an AI agent to screen job applicants for a healthcare business; Memvid itself is not sold as a hiring tool. The CEO's name is Mohamed Mohamed on the company's pages and Mohamed Omar in the Teknovation article.

Sources

The grade next to each source is the best grade that the source can support. "Read in full" means that the text of the page was downloaded and the numbers were checked in it.

  1. D THIRD dealflow newsletter, 31 July 2026: Every AI Agent Needs Memory: 40K Developers, $20K ARR, $800K Pre-Seed
    Read in full in a copy downloaded by the pipeline on 7 November 2026. A newsletter for investors in startups that are raising money. Only source for the revenue figure, the target and the round.
  2. C Teknovation.biz, 28 June 2026, by Katelyn Biefeldt: The founders of Memvid are trying to fix AI's worst habit: forgetting
    Read in full. A regional news interview with the CEO. Source for the founders' background, the investors, the base in Chattanooga, the developer count and the "AI Bully" campaign.
  3. C Memvid: About Us, Our Story
    Read in full. Source for the founders' full names and roles, the origin and the prototype's reach.
  4. C Memvid documentation: Capacity & Plans
    Read the plan table. Source for the prices and limits.
  5. C Memvid: Frequently Asked Questions
    Read in full. Source for the free tier, the paid-plan file limit, the open-source license and the SDKs.
  6. B GitHub: memvid/memvid repository
    Read the repository page and README on 7 November 2026. Source for the stars, forks, commits, license, embedding model and SDKs.
  7. C Hacker News, 5 January 2026: Show HN: I accidentally built "SQLite for AI memory" (Memvid)
    Read in full. Post by Saleban Olow (user salebanolow). Source for the prototype's reach and the Smart Frame design.
  8. C Memvid blog, 4 January 2026, by Saleban Olow: Introducing Memvid V2: Portable, Deterministic Memory for AI
    Read in full. Source for the V2 release date, the Rust rewrite and the file layout.
  9. C Memvid home page
    Read in full on 7 November 2026. Source for the logo wall, the testimonials and the comparison with vector databases.
  10. C Memvid: enterprise product page
    Read the sections on deployment and support. Source for the end-to-end deployment by the company's team.
  11. C Memvid: Careers
    Read in full. Says "a small, focused team" with no headcount.
  12. C Database of failed startups, local export
    Local export. AI-written summaries, not independently verified. Source for the failed predecessors.

Run this company? Send corrections or a verified figure to aiolum@livyden.com

Back to all stories