A person at a table with four different stacks of documents, comparing options

The AI memory market in 2026: four categories, how each one charges, and who each fits

Market9 min read

Agent memory layers, assistant memory, enterprise search and DIY frameworks: public pricing checked on 2026-11-14 and which option fits which kind of team.

by Equipo MemoryFirst

Table of contents
  1. The four categories
  2. How each one charges, using only what they publish
  3. Who each category suits
  4. Where we stand
  5. How to choose in five questions
  6. One-to-one comparisons
  7. What we could not verify
  8. Sources

If you searched for "AI memory" this year, you will have seen products that look nothing alike: an API that lets your agent remember its users, a toggle in your assistant that saves your preferences, a search engine that indexes the whole company, and a pile of libraries to build it yourself. They all use the word memory. They solve different problems and are paid for in very different ways.

This map sorts the market into categories, explains how each one charges using public information from their websites checked on 2026-11-14, and says who each option suits. We are an interested party: we build MemoryFirst. So you will see what we verified and what we did not.

The four categories

CategoryExamplesWhat it remembersWho uses it
Memory layer for agentsMem0, Zep, Letta, SupermemoryWhat an agent you build needs to know about its users or tasksTeams developing products with agents
Memory built into assistantsChatGPT, ClaudeEach person's preferences and context inside that assistantAnyone using that assistant
Enterprise searchGleanThe company's documents and apps, with permissionsLarge companies with many tools
Frameworks and in-house partsOpen-source libraries, Postgres with vectorsWhatever you programTeams with engineering to spare

MemoryFirst sits between the first and the third: it is a memory layer your agents consume over MCP, but aimed at a team's or company's knowledge (decisions, customers, deadlines), with citations. More on that below.

Four filing cabinet drawers open at different heights

How each one charges, using only what they publish

Plans and prices change fast. Everything below comes from the public pages cited under "Sources", checked on 2026-11-14. Where a detail did not appear, we say so.

ProductWhat its public site showsWhat is not visible
Mem0A Free plan (10,000 adds and 1,000 retrievals a month for the first 30 days; 50,000 and 10,000 with your own LLM key) and a Custom plan with price on request, adding graph memory, SSO, audit logs, SLA and on-prem deployment. An open-source library under Apache 2.0Intermediate prices. Our July comparison listed $19 and $249 tiers that do not appear on the page today: verify the current ones
ZepFree with 10,000 credits a month; Flex at $125 a month (50,000 credits); Flex Plus at $375 a month (200,000); custom Enterprise with SOC 2 Type II, a HIPAA BAA and options to deploy with your own keys or in your cloudAnnual price (the page mentions 17% savings with no figures). Graphiti, its graph engine, is open source under Apache-2.0
LettaFree at $0 (up to 3 stateful agents), Pro at $20 a month with $20 in credits, Teams Pro at $20 per seat and custom EnterpriseLetta's pricing pages change often and gave us different versions: we could not verify other plans that appear on third-party sites
SupermemoryFree ($5 in monthly credits), Pro at $19 a month (3 seats), Max at $100, Scale at $399 (SOC 2, HIPAA BAA, self-hosted option) and custom Enterprise. The engine is not open source; the main repository, the MCP server and the plugins are (MIT)Explicit usage limits per plan
GleanPublishes no prices: the button is "Get a demo"Any figure. The per-user estimate that circulates is one we could not check
ChatGPT and ClaudeMemory inside each assistant, managed by the person. In Claude, each project has its own memory and, on Team and Enterprise, the owner switches the feature on or off without being able to see members' memoriesWhat each ChatGPT plan includes: OpenAI's help page could not be opened, and we claim no more than what we saw in search results
FrameworksLibraries and databases under open licences. The cost is yours: engineering and operationsWe did not verify prices or licences of specific frameworks beyond Mem0 and Graphiti

Two useful readings. First: the developer layers charge by usage (credits, adds, retrievals), and the cost grows with your product. Second: what a plan really costs often sits in the security fine print (SOC 2, HIPAA, deployment in your cloud), reserved almost always for the custom or most expensive plan.

Who each category suits

Agent layers (Mem0, Zep, Letta, Supermemory). If you are building a product with agents and each user needs memory, this is your category. Zep stands out for temporal graphs with an SDK; Mem0 publishes its library under an open licence alongside the managed service; Letta offers stateful agents; Supermemory offers plugins for Claude Code and Cursor and a $19 entry plan. It is the best fit when you want to control how it integrates into your code.

Assistant memory. If you work alone, with one assistant, and it is enough that it remembers your preferences, you already have it. The limits: it lives inside that assistant (what Claude knows, your agent in Cursor does not) and it belongs to one person, not a team.

Enterprise search (Glean). If your problem is that hundreds of people cannot find documents scattered across dozens of applications, with per-user permissions, this is the right category. Glean says it has more than 275 connectors and permission-enforced search. To see a price you have to talk to sales.

Frameworks. If your team has engineering to spare and memory is part of your competitive edge, building it makes sense. If not, you are paying in hours what others charge as a subscription, and maintaining a memory system (quotas, deletion, audit, isolation) costs more than it looks on the first afternoon.

A small team debating options in front of a glass wall with sticky notes

Where we stand

MemoryFirst is a managed memory for a team's knowledge, consumed by your agents and assistants over MCP (Claude, ChatGPT, Cursor and VS Code), with hooks for Claude Code. Every answer cites its source, and contradicting facts resolve toward the most recent one.

  • Plans: Free (one project, 30-day retention, one seat), Pro at €12 a month billed annually (€14.40 month to month) with unlimited projects, and Team at €39 a month billed annually, five seats and shared memory. There is also a Pioneer offer at €9 a month until 2026-12-31 for the first 50 seats. VAT included; each extra Team seat is €9 a month billed annually. See the plans.
  • Data: database, embeddings and original files in the EU. Default inference is not EU-only: the question and the passages go to a provider in the US, unless you bring your own key.
  • What we do not have: a self-hosted edition, an audio voice channel (only transcripts you bring), and isolation by physical database (planned, not in production).

Where we differ: we think about teams that use ready-made agents, not people building an agent, and our answer includes the citation. Where we lose: if you need a deep SDK, the most flexible temporal graphs, deployment in your cloud or published certifications, the other options are more mature on those pieces.

How to choose in five questions

  1. Are you building a product with agents, or using ready-made agents? The first points to developer layers; the second to team memory.
  2. Does the memory belong to one person or several? Assistants cover the personal; shared memory needs something else.
  3. Can you show the source of each answer? If your work is regulated, ask before you ask the price.
  4. Where is your text processed? Ask about the inference provider, not just where the database sits.
  5. What happens when you want to leave? Export and verifiable deletion, or you stay tied.
An independent developer working in the morning at a laptop

One-to-one comparisons

We wrote each comparison separately, with pros and cons of both sides and a "when to choose the other" section:

For the reasoning behind all this, read why AI agents need a memory layer. And to see it applied, look at the software team use case and the one for an agency with many clients, both simulated scenarios.

What we could not verify

  • Mem0 plans and prices other than Free and Custom.
  • Letta plans that appear only on third-party sites, and whether its API platform has a separate price.
  • Glean's price and any per-user figure.
  • What each ChatGPT plan includes regarding memory. For Claude we only state what its official help page says; third-party reports about which plan has which version we could not cross-check.
  • Prices and licences of specific open-source frameworks.

Sources

All accessed 2026-11-14.

  • Mem0, pricing: https://mem0.ai/pricing (accessed 2026-11-14)
  • Mem0, repository and licence: https://github.com/mem0ai/mem0 (accessed 2026-11-14)
  • Mem0, platform description: https://docs.mem0.ai/platform/overview (accessed 2026-11-14)
  • Zep, pricing: https://www.getzep.com/pricing (accessed 2026-11-14)
  • Zep, home page: https://www.getzep.com/ (accessed 2026-11-14)
  • Graphiti, repository and licence: https://github.com/getzep/graphiti (accessed 2026-11-14)
  • Letta, pricing (redirects to the docs): https://docs.letta.com/letta-code/pricing (accessed 2026-11-14)
  • Letta, API plans: https://docs.letta.com/guides/api/plans (accessed 2026-11-14)
  • Supermemory, pricing: https://supermemory.ai/pricing (accessed 2026-11-14)
  • Supermemory, home page and licences: https://supermemory.ai/ (accessed 2026-11-14)
  • Glean, pricing: https://www.glean.com/pricing (accessed 2026-11-14)
  • Glean, home page and product: https://www.glean.com/ and https://www.glean.com/product/assistant (accessed 2026-11-14)
  • Claude, how memory works: https://support.claude.com/en/articles/11817273-how-does-claude-s-memory-work (accessed 2026-11-14)
  • ChatGPT, memory (search result; could not be opened directly): https://help.openai.com/en/articles/8590148-memory-and-new-chat-controls-for-chatgpt (accessed 2026-11-14)

See where MemoryFirst fits

Start free with one project and compare it against whatever you use today, with your own data.

Start free

Keep reading