What Mem0’s own docs say about lock-in

2026. Framework-agnostic isn’t a nice-to-have — it’s the adoption bar.

Agent-memory vendors are fond of telling you about integration counts. Mem0’s own positioning boasts dozens of LLM frameworks and vector stores supported. Read closely and the message is: a memory layer that locks you to one framework or one vector store will not be adopted at scale. That’s a real insight — and it’s one we agree with, and act on in a way that doesn’t create a different lock-in.

The two kinds of lock-in

1. Framework lock-in: “this memory only works inside my agent SDK.” Adopt it and your memory is hostage to a framework choice you may reverse later.
2. Service lock-in: “your memory lives in my datacenter.” Adopt it and your data — and your recall latency, and your bill — is hostage to a vendor’s uptime, pricing, and compliance posture.

We avoid both, not by advertising more integrations, but by refusing to define memory through a proprietary channel at all.

Brain Server’s no-lock-in answer

• UMP 1.0 / L3 conformance — a published memory-protocol standard, scored by the reference conformance suite (13/13, L3). Your memory is readable and writable through a standard wire, not a private API. Leave our product and the protocol — and your data — travel with you.
• An open HTTP contract — GET /openapi.yaml documents every route, served by the binary itself. Any client, any language, no SDK required.
• MCP — a stateless core implementing the Model Context Protocol, so it slots into the agent tools ecosystem without being bound to one runtime.
• Local-first storage — a single SQLite-family file on your device. There is no cloud side, no egress, no “your memory in our cluster.” The ultimate anti-lock-in is that there’s nothing to be locked into.

The honest trade

No vendor lock-in means no vendor magic. The deterministic, no-LLM retrieval is yours to run — which also means the curation and evaluation are yours too (the corpus-quality ceiling in the PPR explainer is a real operator step, not a marketing asterisk). We think that’s the right trade: portability and audit over convenience. A memory store you can leave is a memory store you can trust; a memory store you can’t leave is a dependency you’ll be stuck defending.

The takeaway: when you evaluate agent memory, don’t count integrations — count standards. Ask: is there a published protocol? An open contract? A local file I own? Those are the things that survive a framework migration, a vendor pricing change, or a compliance deadline. That’s what “no lock-in” actually means, and it’s the bar we hold ourselves to.

See docs/trust/proof-map.md for the UMP L3 + capability-token rows and how to verify them live.