Supermemory Pricing in 2026: What It Really Costs

If you are trying to work out supermemory pricing before committing an engineering sprint to it, the honest starting point is that the number on the pricing page is not the number you will pay. Supermemory runs a hybrid model: a flat monthly fee that includes a bundle of usage, then metered rates for whatever you use beyond it. That is a sensible way to price a product where one customer stores a handful of short notes and another ingests a warehouse of PDFs, but it means the headline "$19 a month" answers a different question than "what will this cost my team."

This review covers what each tier actually includes, what the metered rates mean in practice, whether the "open source" claim you will see repeated across comparison sites holds up, and how the whole structure stacks up against Mem0, the tool Supermemory is most often pitched against.

What does Supermemory pricing actually look like in 2026?

As of September 2026, Supermemory's public pricing page lists five tiers. Four of them have a fixed monthly price; the fifth is negotiated.

Table 1: Supermemory's published plans and included usage credit, from Supermemory's pricing page, September 2026.

Underneath every tier sits the same metered rate card, and this is the part a plan comparison alone will not show you: storing plain text runs $5 per million tokens, storing richer content (files, images, anything Supermemory calls "rich content") runs $10 per million tokens, and Supermemory's own retrieval layer, which it calls SuperRAG, bills separately again at $1 to $2 per million tokens. Search and graph traversal cost $5 per million queries, and background operations run $100 per million. Supermemory says only unique content is billed, with deduplication happening at the token level, so re-ingesting something you already stored should not double the charge.

We ran a rough sense check on what that metering means for a mid-sized workload: a team ingesting around 4 million tokens of plain-text notes a month and running 400,000 searches would spend roughly $20 on storage and $2 on search, which is $22 in raw usage. That alone clears the Pro plan's $20 credit, before counting a single team seat or connector. It is not a huge overage, but it is the kind of thing that quietly nudges a team from Pro toward Max well before the seat limit does, and it is worth modelling before you pick a tier rather than after the first invoice.

Is Supermemory open source, or is that a marketing claim?

This is one of the more genuinely good answers in the agent memory space right now, and it is worth stating plainly because a lot of "AI memory" tools use "open source" loosely to mean "you can read some of the code."

Supermemory's core memory and context engine is published on GitHub under the MIT licence, one of the most permissive licences going, and its own self-hosting documentation describes the self-hosted deployment as running from a single binary: no Docker, no database to provision beforehand, no separate config step. It boots with an embedded graph engine and built-in local embeddings, and it uses your own model key for the actual extraction step rather than a hosted one. According to Supermemory's self-hosting overview, there is no stated usage limit on that self-hosted path, which makes it a real self hosted ai memory option rather than a locked demo pointed at a sign-up form.

What is not open, and this is the distinction worth holding onto, is the managed cloud service sitting on top of that engine. The proprietary extraction models Supermemory describes as higher quality and cheaper at scale, the hosted connectors for Gmail, Google Drive, OneDrive and S3, and the SOC 2/HIPAA-backed infrastructure on the Scale and Enterprise tiers are commercial features layered on an open core. That is a completely standard open-core structure, not a bait and switch, but it does mean "is supermemory open source" has a two-part answer: the engine, yes; the full managed product, no.

Gartner projects that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025.
Gartner newsroom, August 2025

That prediction is worth keeping in view here, because it is the reason agent memory pricing has become a real budget line rather than a curiosity. If a meaningful share of enterprise software is about to carry a task-specific agent, the tool that agent uses to remember things stops being an experiment and starts being infrastructure someone has to plan a line item for.

How does Supermemory's pricing compare to Mem0's?

Mem0 is the tool most people researching supermemory pricing are also comparing against, since both sit in the agent memory layer and both are frequently searched together as alternatives to each other. The two companies structure their pricing around different units, which makes a naive dollar-for-dollar comparison misleading.

Table 2: entry-level pricing units for Supermemory and Mem0, both as published in September 2026.

The practical difference: Supermemory's bill tracks how much content you push through it, measured in tokens, so a team ingesting long documents or rich media pays more even if they rarely search. Mem0's bill tracks how many discrete memory operations happen, so a team with lots of short facts and frequent lookups pays more there even on modest token volume. Neither structure is better in the abstract. Which one is cheaper for you depends entirely on whether your workload is content-heavy or operation-heavy, and the only reliable way to answer that is to estimate your actual monthly token and call counts against both rate cards before committing, not after.

Mem0's core engine is also open source and MIT licensed, much like Supermemory's, so "closed source ai memory" is not really the axis these two compete on. Where they differ more sharply is what each one is built to remember in the first place: Mem0 extracts durable facts from conversations a specific user has with an assistant, while Supermemory positions itself more broadly as a general memory API with connectors into a user's own documents and inbox. If what you actually need is neither: a shared, permission-aware view of a company's own systems of record rather than one person's conversational history or personal document memory, that is a different problem again, and it is the one Contextely is built around. We are honest that our own retrieval is lexical by default, meaning it finds a stored memory by matching terms rather than meaning unless an operator explicitly configures an embedding model, which is a real trade-off against tools that always do semantic search out of the box.

What actually drives the bill, tier by tier

A few patterns are worth flagging separately from the raw numbers, because they are the parts a quick skim of the pricing page tends to miss.

• The jump from Pro to Max is really about the Gmail connector and a 6x credit multiplier, not new API capability. If you do not need inbox ingestion, the extra credit alone has to justify $81 a month.
• Scale's SOC 2 and HIPAA support matters far more to regulated buyers than its raw credit bump. If compliance paperwork is the blocker, that is the tier to budget for regardless of usage volume.
• The metered rates apply even inside a paid plan once you exceed the bundled credit, so a "flat fee" plan is really a flat-fee floor with variable costs stacked on top. Model your expected token and query volume against the rate card in Table 1 before assuming the monthly price is the ceiling.
• Enterprise's air-gapped self-hosting is a genuinely different offer from the free binary: it comes with contractual support and a dedicated account manager, which is what regulated or on-device ai memory deployments are usually paying for, not extra features in the software itself.

None of this makes Supermemory expensive or cheap in isolation. It makes it a metered product wearing a subscription's clothing, which is common in this category and worth pricing out against your own volume rather than the sticker number alone.

Where this leaves a buyer comparing agent memory pricing

If your use case is a single user's assistant that needs to remember preferences or personal documents, and you are comfortable running your own binary and model key, Supermemory's free self-hosted path is a legitimate answer to "local binary memory" with no subscription attached at all. If you need managed infrastructure, hosted connectors or compliance paperwork, the paid tiers are priced in line with comparable memory api pricing elsewhere in the category, and the metered add-ons are the part to model carefully rather than the monthly fee.

If what you are actually solving is company-wide retrieval with per-person entitlement, rather than one user's personal memory, that is a narrower and different problem than either Supermemory or Mem0 is built for, and it is worth reading how Mem0 and Contextely differ or the broader llm memory database comparison before assuming a personal-memory tool is the right fit for a company-wide deployment. For teams weighing whether to run any of this themselves, the self-hosted AI memory setup guide and Contextely's own self-host and pricing pages are a reasonable next stop for comparison.