Mem0 wants AI to remember you. The trick is knowing what to forget.

In December 2023, a chatbot built around the teachings of Sadhguru acquired a rather worldly problem. People liked talking to it. Then they returned, expecting the conversation to have meant something. The bot could discuss a spiritual journey; it could not remember the person taking one.

Taranjeet Singh later recalled the complaint in a public founder account: “This is cool, but it doesn't remember anything about my meditative journey.” He and Deshraj Yadav had been building Embedchain, an open-source framework for giving AI access to external information. Their experiment exposed a different need. Finding the right teaching was one task. Remembering what a particular person had already said was another.

01 / The inconvenience hiding in plain sight

There is a small indignity in having to introduce yourself to the same software twice. For a novelty chatbot, it may be tolerable. For a tutor, it wastes the next lesson. For customer support, it turns yesterday's explanation into today's homework. A conversational interface makes an implicit promise of continuity. A fresh session can promptly break it.

Mem0 occupies the space between those expectations and the machinery. A language model can use the information supplied in its current context; lasting knowledge about an individual requires an application to preserve and reintroduce it. Mem0 gives developers an engine for that work. It does not replace the model. It decides which pieces of past interaction deserve another appearance.

The distinction matters. A transcript records everything said, including the irrelevant bits. A memory system must extract a preference, recognize when it changes, and find it when it becomes useful. Consider an illustrative assistant that first hears “I live in London” and later “I've moved to Lisbon.” Two faithfully stored sentences are easy. A useful answer about nearby restaurants requires handling the change.

02 / A very short API, a rather long job

The managed platform accepts conversations, distills facts and returns relevant memories at query time. Its SDKs let an application attach those operations to a user or agent. Developers can retrieve context before answering and save durable information afterward. The simplicity lives at the interface; extraction, deduplication and ranking live behind it.

Mem0 also distributes an Apache-licensed open-source engine. Teams can run a library in their application or a self-hosted server, choosing the models, embeddings and storage underneath. The original company description presented a hybrid of vector, graph and key-value storage: meaning, relationships and structured facts each have a job. The current open-source defaults differ by deployment. Graph memory should not be assumed to accompany every installation.

This is where Mem0 fits in the market: infrastructure for builders, rather than another general-purpose assistant. A database can hold the information, but the builder still has to decide what to extract and retrieve. Framework memory tools and agent platforms offer other routes. Mem0's offer is to take enough of that work off the engineer's desk to justify a separate dependency.

Their backgrounds help explain the combination. Singh worked on growth and product at Khatabook after starting his engineering career at Paytm. Yadav built machine-learning infrastructure at Tesla Autopilot. They had met in college and worked together on open source, including EvalAI. One had experience getting software into people's hands; the other had experience making demanding systems operate.

03 / The algorithm could wait. The users could not.

Even a company founded on remembering users can temporarily forget to consult them. Singh says the team spent two months refining its memory algorithm without shipping anything visible. At a YC retreat, a partner asked why they had not launched. Within 36 hours, according to his account, they did.

The cost he identifies is time: weeks spent perfecting an internal answer before customers could supply the next question. The useful lesson is quite specific. Once there is a working product and evidence of demand, a release creates information that another private round of polishing cannot. A team building memory particularly needs to see what users actually expect it to retain.

On October 28, 2025, Mem0 announced approximately $24 million across seed and Series A financing. Kindred Ventures led the seed; Basis Set Ventures led the Series A, with other participants including Peak XV Partners, GitHub Fund and Y Combinator. Mem0 reported that quarterly API calls had risen from 35 million in Q1 to 186 million in Q3 2025. Those are operations, not revenue or paying accounts, but they show substantially more work passing through the system.

04 / The tutor remembers where you got stuck

OpenNote offers a useful customer example because learning rarely follows a tidy transcript. A student can explore a diagram, leave, and return to a related problem days later. Mem0's published case study describes OpenNote using memory to maintain evolving learner profiles in its Feynman-2 tutoring engine. It reports integration in two days and a 40% reduction in token usage per prompt.

Sunflower Sober presents a different kind of continuity. Its recovery companion needs earlier context to make later check-ins relevant. In Mem0's case study, the team describes previously sending entire histories into prompts. It reports a one-day integration and roughly 70-80% lower token usage. These are customer claims published by the vendor, and they describe engineering outcomes rather than evidence of clinical effectiveness.

“We were throwing the entire memory into context - not scalable.”

Sunflower engineering lead, in Mem0's customer account

Both examples expose the same practical choice. Teams could build summarization, storage and retrieval themselves. Buying memory lets them spend those hours on the tutor or companion instead. The decision makes sense when repeated interactions are central to the product and the retrieved information improves the next response. An occasional, self-contained question provides much less reason to maintain a history.

05 / Memory has a bill of its own

Mem0's managed business starts with a free Hobby tier. At the time of writing, Starter costs $19 a month and Pro $249, with custom enterprise pricing. Add requests and retrieval requests are counted separately. A cheap subscription is not the entire calculation: workload, request limits and the rest of the application's model usage still matter. Open source removes the software subscription, while leaving infrastructure and configured model calls to the operator.

The company's April 2025 research reported a 26% relative improvement over its OpenAI memory baseline on LoCoMo. Against processing full histories, it reported 91% lower p95 latency and more than 90% token-cost savings. These are comparisons within a particular experiment, with particular baselines. They are a reason to test the approach, not a promise about every application.

A newer development sharpens the question. Mem0 released DolphinBench in September 2026 to test actions in simulated applications, alongside cost and latency. Remembering a channel rule must lead to posting in the correct channel. The published results vary by model and agent setup; some configurations spend more after adding memory. The honest buying question is how much the system costs to complete the tasks you need it to complete.

06 / Take the relationship with you

Mem0's distinctive argument is portability. Native memory inside a model provider's product may suit someone staying there. A developer switching models, or distributing context among several agents, has a different concern: accumulated user knowledge should remain available. OpenMemory's local-first MCP server and browser extension extend that idea to compatible tools, with interfaces for inspecting and managing memories.

Distribution is getting easier too. In September 2026, Mem0 announced a Vercel Marketplace integration with project provisioning and billing through Vercel. AWS Strands documentation also includes a Mem0 memory-agent example. These integrations put memory nearer the places developers already build. They do not remove the need to scope information correctly or check that a changed preference actually replaces an obsolete one.

For a builder, the experiment to copy is modest. Use a stable user identity. Save a durable preference. Retrieve it in a later session. Change it, and check what the assistant says next. Measure completed tasks, response time and total spend against the existing approach. That test will tell you more than a long chat that merely looks convincing.

Mem0 began with people asking software to remember their journey. Its business depends on making that expectation affordable and dependable for someone else to implement. The charm of the idea is also its discipline: an assistant should remember enough to spare you an explanation, and update enough to spare you an argument.