Roadmap: opt-in, human-reviewed memory sharing for agent packs

[lodar]

Let a published agent pack optionally carry a redacted, human-reviewed slice of team memory, so an imported agent boots already knowing the company. Never automatic — always opt-in and reviewed. Feeds the pack export/import pipeline.

[hegu-1]

The opt-in + human-review boundary is the right starting point. I would make the exported object a reviewed projection, not a slice of the live memory store. That keeps three things separate:

1. source memory remains private and mutable;
2. the pack contains an immutable snapshot with provenance;
3. import creates candidates, not immediately active memory.

A small manifest could carry source_ids, redaction_version, reviewed_by, reviewed_at, scope, and a content hash. On import, the receiving user should see a diff and choose which entries become active. That matters because even a correctly redacted memory can be stale, organization-specific, or inappropriate for the receiving agent's authority.

I would also keep deletion semantics asymmetric: removing an imported pack should remove its projection and derived index entries without touching any locally authored memory. A round-trip test worth adding is export -> import -> reject one entry -> re-export; the rejected entry must not silently return through provenance expansion.

This is close to the source/snapshot boundary I use in Personal AI OS, where continuity-sensitive data stays user-controlled and exportable rather than becoming provider-owned state: https://github.com/hegu-1/personal-ai-os/blob/main/docs/privacy.md

[lodar]

This is the model we're building toward: a reviewed projection over a live slice, a provenance-stamped immutable snapshot, and import as candidates with a human diff before anything activates. Deny-default scoping is the substrate underneath.

Adopting your export -> import -> reject -> re-export round-trip as an acceptance test: the rejected entry must not return through provenance expansion. That invariant wasn't written down; now it will be.

Asymmetric delete is right too, drop the imported projection plus derived index entries, never touch locally-authored memory. Thanks for the personal-ai-os pointer.

[lodar]

Progress: scoped memory inheritance shipped as agent create --inherit-memory=<scope>, with deny-default scoping as the substrate underneath.

The reviewed-projection model above is the next build on top: an immutable, provenance-stamped snapshot carried in the pack (not a slice of the live store), import as candidates with a human diff before anything activates, and the export -> import -> reject -> re-export round-trip as an acceptance test where a rejected entry must not return through provenance expansion. Always opt-in and human-reviewed, never automatic.