wrt our discussion in https://github.com/GetStream/Vision-Agents/issues/646

Interesting idea about integrating a durable memory adapter. I'd look into how current memory handling works with Stream Chat to see how this could fit in.

Hi @alisaifee, @tbarbugli I could raise a PR for this to get started with, happy to build this end to end rather than hand you a design you'd have to maintain.

a package under plugins/, following the same pattern as the existing vendor plugins installed separately, opt-in, inert unless configured. Nothing in core needs to know MemCode exists, and if you later build a first-party memory path, the plugin implements the same interface or simply gets deleted.

Surfaces, whichever you prefer:

• Plugin package (our recommendation), Python, matching the existing plugin pattern, maintained by us.
• MCP server, if you'd rather memory be tools on a transport.
• Direct SDK: Python, TS or Go clients, if you want it in-process behind your own interface.

We'll do the whole thing: the plugin, the Stream Chat sync, docs, and an example agent that remembers a user across calls. Scoping is per-user and per-channel, so nothing crosses between them.

Self-hosted: if you'd want this running on your own infrastructure rather than against our API, that's available, worth its own conversation once the shape is settled.