30 days: Mem0 Platform kills self-serve plans and makes Free users bring the LLM key
Mem0 Platform removed self-serve Starter, Growth and Pro. Free users get Mem0-paid model calls for 30 days, then writes need your own key.
Kill posts8 min read
Tim Neunzig
30 days is the new grace period before Mem0 Platform stops paying for your model calls on Free.
Mem0 moved the hosted Platform from visible self-serve tiers to Free plus Custom. Starter, Growth and Pro are closed to new accounts. If you cancel a legacy plan, you cannot restart it.
That is not a pricing tweak. That is a dependency lesson. Your agent memory write path now sits behind a plan page that can disappear, a sales conversation, or your own OpenAI, Anthropic, or OpenRouter bill.
What happened
Around 2026-10-26/27, Mem0 removed transparent self-serve paid pricing for the hosted Mem0 Platform. The current shape is Free and Custom. The old Starter, Growth and Pro plans are legacy. New accounts cannot buy them. Cancelled legacy plans cannot be restarted. Plan changes go through the Mem0 team.
The old numbers still matter because they show the rent you were expected to normalize. PricingSaaS reported the former Starter plan at $19 per month and the former Pro plan at $249 per month. The kill sheet uses the Pro legacy tier for the rent math: $249 flat per month for 36 months is $8,964.
The Free plan is not just free usage. It is a timed handoff. The pricing docs describe the first period as 30 days with Mem0-paid model calls. The quoted allowance is "10,000 adds / 1,000 retrievals" in that first 30-day window. After that, Free switches to bring your own key. The later allowance is listed as "50,000 adds / 10,000 retrievals", but your OpenAI, Anthropic, or OpenRouter provider bills the model calls.
So the daily failure mode is simple. Your app keeps calling the memory add endpoint. The 30 days end. If you did not connect your own provider key, new memory adds fail. The agent still answers, but its long-term memory stops learning. That is worse than a clean outage. It rots quietly.
Mem0 is also not a dead open-source project. The kill sheet lists Mem0 open source at 68,321 GitHub stars and a last commit date of 2026-10-30. The open-source library is Apache-2.0. That matters. The hosted plan changed. The underlying pattern can still be owned.
Mem0 says 150,000+ developers build with it. That is the point. A tool used inside agent infrastructure is not a nice dashboard you can swap out on Friday afternoon. It sits in the path between a user saying something useful and your product remembering it.
What it actually did for you
You used Mem0 Platform for one core job: durable memory for agents and AI applications.
Daily use was not mysterious. Your app sent conversation snippets to an API. The service extracted stable facts. It stored them against a user, an agent, or an organisation. Later, your agent asked for relevant memories and got back context instead of replaying a full transcript.
That is the useful part.
The daily features were these:
- Add memories from conversations. This is the write path. Without it, the agent forgets new facts.
- Semantic memory search and retrieval. This is the read path. The agent needs relevant old facts at response time.
- Per-user and per-agent scoping. Memory cannot leak between users, tenants, or agents.
- Bring your own LLM key. On Free, this becomes mandatory after 30 days if you want new memory adds to keep working.
- Basic usage and failure visibility. You need to see adds, retrievals, blocked writes, and key failures before customers notice the agent went stupid.
Everything else is where rent fattens.
Graph memory on Custom may be useful for some teams. Unknown for yours unless your agent actually queries graph relationships. Dream consolidation sounds expensive to operate and easy to postpone. On-prem deployment as a vendor-managed feature is still vendor-managed. SSO and audit logs matter in certain companies, but they do not make memory extraction better. Vercel Marketplace packaging is packaging. Full MCP server integration is noise if your product only needs a direct internal API.
The part your users feel is much smaller than the menu.
They do not care which plan name stores a memory. They care that the agent remembers their preferences, facts, and prior instructions without mixing them with another account. They care that memory writes do not stop because a free window ended and no one connected a key.
The math
| Tier | Price per user per month | Team size | 36 months | Total |
|---|---|---|---|---|
| Pro, legacy, closed to new accounts | Unknown per user, $249 flat/month | 10 | 36 | $8,964 |
| Owned build range | n/a | 10 | one build, not rent | $2,500 to $8,000 |
You keep up to $964 of the 3-year Pro rent after funding the high end of the owned build, before hosting and LLM provider costs, which are separate and unknown.
At the $8,000 high end, payback is 33 months against the old $249 per month Pro rent. At the $2,500 low end, the simple payback is sooner, but the kill sheet gives the official payback number as 33 months.
This table is deliberately boring. It does not pretend Custom has a price, because Custom has no public list price in the kill sheet. It does not invent a per-seat number, because the old Pro figure is a flat $249 per month. It does not count provider model calls, because those depend on your usage and provider.
The key fact is enough. Three years of the old Pro plan is $8,964. The owned replacement range is $2,500 to $8,000. If memory is core infrastructure, that spread pays for control.
What you would build instead
You do not rebuild a platform brochure. You build the memory path your agents use.
- Memory write API. A small service receives conversation snippets, calls your configured LLM provider, extracts durable memories, and stores them with user_id and agent_id scopes.
- Retrieval API. The service runs semantic search over stored memories and returns relevant context to the agent backend.
- Tenant and scope rules. Organisation, user, and agent identifiers are mandatory on writes and reads. Cross-tenant memory leakage is blocked by design.
- Provider key manager. You store the team’s OpenAI, Anthropic, or OpenRouter key and use it for extraction calls. This is not an afterthought after 30 days. It is day-one plumbing.
- Usage and failure dashboard. A plain web UI shows adds, retrievals, provider key status, and failed or blocked writes.
- Data migration. If you still have application-side conversation history, replay or batch it into the new memory store using existing user and agent IDs. If not, start empty and populate from new conversations.
The stack is small. Use the Mem0 open-source library for the memory layer. Use Postgres for users, agents, memories, usage records, and provider-key metadata. Add pgvector for semantic search inside Postgres. Put a small FastAPI service in front of it. Run it with Docker Compose on one self-hosted server.
Open-source building blocks:
- Mem0 open source, Apache-2.0, core memory extraction, storage, and retrieval.
- Postgres, PostgreSQL License, primary database.
- pgvector, PostgreSQL License, vector search in Postgres.
- FastAPI, MIT, small Python API for add and retrieve endpoints.
Hosting is one self-hosted server with Docker Compose. The glue code is yours. The memory data sits under your backup, retention, and deletion rules. LLM provider costs stay separate, because they were always real costs. The difference is that your write path is not waiting for a hosted plan rule to change.
What you lose is also real. You lose Mem0-hosted operations, updates, and scaling. You lose transparent access to the old Starter, Growth, and Pro plans if you cancel them, because the old plan cannot be restarted. You lose Mem0-paid model calls on Free after the first 30 days. You lose Custom-tier vendor features such as graph memory, dream consolidation, vendor-managed SSO, and audit logs. You also do not reproduce the Vercel Marketplace organisation path, because the kill sheet says Vercel Marketplace organisations cannot use BYOK.
That is fine if your actual product needs memory add, memory retrieve, scoping, a key manager, and basic monitoring. It is not fine if you bought Mem0 for every advanced feature on the menu.
Who should do this
Build this if you run a chatbot or agent product where memory quality is part of the user experience. If the agent forgets preferences, prior facts, or account context, users notice. You should not let that depend on whether a 30-day provider-paid window expired.
Build this if you are an AI startup with multiple tenants and a predictable memory workflow. You need add, retrieve, user scope, agent scope, usage visibility, and provider key control. You do not need graph memory, dream consolidation, SSO, or marketplace packaging to ship the thing your users touch.
Build this if you were on, or were planning around, the old Pro-style economics. $8,964 over 36 months is enough to fund a narrow owned service in the $2,500 to $8,000 range. The payback is 33 months at the high end, and you stop treating agent memory as rented plumbing.
Keep renting if your team has no one to operate Postgres, Docker Compose, backups, and provider keys. Keep renting if you need Mem0-hosted scaling more than you need control. Keep renting if Custom-only features are the actual reason you use the product. Keep renting if your usage is experimental and the Free plan, with your own model key after 30 days, is enough.
Do not build because you hate a pricing page. Build because the hosted product sits on a critical write path and the rent buys features you do not use.
The SaaSKiller Take If memory is core infrastructure for your agents, renting the write path from a vendor that can remove self-serve tiers is a bad dependency. A legacy Pro plan is $8,964 over three years before your own LLM costs. Build the narrow memory service you actually use and keep the data on your server.
Sources
- https://docs.mem0.ai/platform/pricing
- https://docs.mem0.ai/platform/pricing.md
- https://mem0.ai/pricing
- https://mem0.ai
- https://pricingsaas.com/news/article/mem0-drops-starter-and-pro-self-serve-pricing-for-free-and-custom-plans
AI-generated