in reviewhumanPR #332Identity
Remembrane: redirect handshake for scoped health context over the mailbox and A2A, with adversarial-requester and fail-closed tests
Remembrane is a patient-owned health memory for AI agents: other agents get a purpose-scoped view of a person's
Description
The pitch.
### Summary Remembrane is a patient-owned health memory for AI agents: other agents get a purpose-scoped view of a person's health information, with a source for every item. This PR adds `examples/integrations/remembrane-redirect`. In it, a patient's personal agent never answers health requests itself. It redirects each requester to the memory with a short-lived token scoped to the purpose assigned from the requester's identity card, or declines. ### Integration - **This PR:** the personal agent, a standard-library Town mailbox participant that can also serve A2A with `--a2a`; a check script; synthetic requester fixtures; recorded evidence; and one test. - **Outside Town:** Remembrane runs separately, on loopback, from its own repository. It answers `POST /delegate` (the personal agent gets a token) and `POST /context` (the requester redeems it). - **Cases:** five synthetic requesters (a clinic, a coach, an employer programme, two adversarial coaches) plus an unreachable memory, through Town's mailbox and again over A2A. - **Details:** message formats and flow are in the example's README. ### Reproduce Town `6f9a7c064486c4a0ae9242f1ff341b8a4aad0244`, Remembrane `ad3bb6ab7fa9e807f6e4e3bc8757ca9c136701e2`, Python 3.13. Offline, no credentials: ```sh python -m pip install -e '.[dev]' python examples/integrations/remembrane-redirect/run_check.py --verify examples/integrations/remembrane-redirect/evidence/local.json python -m pytest -q tests/test_example_remembrane_redirect.py ``` The live run and reading what each requester received (`--show`) are described in the README. The Remembrane repository is private until the course's final submission. ### Results and limitations All six cases passed over both the mailbox and A2A, and Town's A2A probe passed: - the clinic got an intake token and the coach a fitness token; every returned item had a source and was in scope; - the employer was declined; - an adversarial coach's claim of a medical role did not widen its scope, and asking for a clinical purpose got it declined; - with the memory down, the personal agent failed closed; - tampered tokens were refused. Untested: - the participants are ours, not independent agents; - identity cards are self-asserted; - scope is checked by category only (no time windows, no wording inside items); - one synthetic patient. The full table and limits are in the README. ### Course context MIT AI Studio, Fall 2026, HW8 submission.
Try it
Open PR on GitHubView diffCheckout locally
git fetch origin pull/332/head:pr-332
git checkout pr-332