feat: Project isolation support for multi-tenant deployments

[jack-arturo]

Feature Request

Enable a single AutoMem backend to serve multiple projects with completely separate memory spaces.

Use Case

Teams or organizations that want to:

• Run one AutoMem instance for multiple projects
• Keep memories isolated between projects
• Reduce infrastructure overhead

Proposed Implementation

Originally proposed in #5 by @andrewleech:

1. Add projectId to AutoMemConfig interface
2. Send X-Project-ID header when project is configured
3. Add --project-id CLI flag to setup command
4. Support AUTOMEM_PROJECT_ID environment variable
5. Load project ID in MCP server

Notes

This requires corresponding backend changes to scope Memory nodes and Qdrant collections by project_id.

---

Tracking issue created from closed PR #5

[George-RD]

Could also consider, a hierarchy? So let's say one person can see/use more projects than others?

User story:

I want to give an AI bot agent its own memories, but there are some I dont want it to access in case it gets hit with prompt injection, so it means it should not be able to access what it isnt allowed.

So propose API key per project.

[jack-arturo]

Right now the default recall prompts use the project name as a tag (like mcp-automem) and this works as a hard gate to limit recall to memories from just that project.

Or for example when recalling memory as the Slack bot, recall is confined to the current workspace by ID unless we specifically ask for a wider recall.

That's worked well for us to keep info compartmentalized but it sometimes means context that is relevant (like from the automem or automem-website) is excluded.

I've been experimenting with letting the tag based recall be looser if there are few / poor matches, but in benchmark tests the quality of the recall has gone down.

At the moment thinking it's best to keep the hard gate on tags. And that does support sharing per-project memories with teammates. Without adding the extra complexity of an API key.

What do you think of that? Can you give me some scenarios where there should be no access whatsoever, vs some intra-team crossover (ie based on a project or code conventions?).