---
status: reconciled-partial
coverage_depth: deep-for-server-backed-core
last_walked: 2026-05-29
last_reconciled: 2026-09-24
specs:
  - apps/oshun/web/e2e/profile-memory-controls.spec.ts
  - apps/oshun/web/e2e/memory-edit-pause-forget.spec.ts
source: WALKTHROUGH/journeys/memory-edit-pause-forget.md
---

# Journey result: Memory review, edit, pause, and forget

- **Walked**: 2026-05-29 against
  `bf12b0f7d8c291499e1df1ec75d4bd21acf3f6b1`. The retained `/tmp` walker is
  ephemeral, so this is a dated smoke record rather than a reproducible run
  artifact. The client-bundle fix later landed in
  `adae8117ddb084fd6dbf9c26b0a31e4f86da2148`; the historical `/v1/memory`
  stub landed in `03d45f0ad7a276369d9536f8f15e7ba0c627d945`.
- **Reconciled**: 2026-07-18 against the current profile component, browser
  reducers, Iris BFF route and durable adapter, startup wiring, and both named
  Playwright specs. This was source review, not a fresh application run.
- **Reconciled again**: 2026-09-24, by source review of
  `8ed8581a6c0` (audit-remediation H13.a: memory clears, opt-outs and the
  full clear reach Iris) and `470ae40e7c1` (H13.b: destructive changes need
  a password step-up the BFF checks and records). Only the full-clear,
  step-up and deletion statements below changed; no suite was rerun.
- **Verdict**: **partial / deep for the server-backed core** — the current
  suites deeply cover Iris hydration, server-authoritative pause, blocked and
  resumed writes, single-record deletion, the seven-day full-clear request, and
  an export receipt. Editing, conflict resolution, and the audit projection
  remain browser-local. Destructive changes now need a password step-up that
  the BFF checks and records on the session.
- **Current authority**:
  [`WALKTHROUGH/journeys/memory-edit-pause-forget.md`](../journeys/memory-edit-pause-forget.md)
  preserves those authority boundaries step by step. Its full-clear and
  step-up steps still describe the behavior before H13.a and H13.b.

## Result at a glance

| Interaction | Proof · limit |
| ----------- | ------------- |
| Iris hydrate/review | **Authority / proof:** `GET /v1/iris/adapter/review?role=settings&pageSize=100` supplies server records; browser state and `oshun-profile-memory-controls-v1` cache the projection. Real-BFF cases cover hydrate, empty, unavailable/retry, consent, filters, detail, and tenant scope. **Limit:** an unavailable read may leave cached rows visible; detail and filtering do not perform a fresh authoritative read. |
| Edit/conflict | **Authority / proof:** local reducers create revisions; tests cover edit, audit projection, telemetry redaction, and two-tab keep-remote/overwrite. **Limit:** no BFF edit, durable revision precondition, or cross-device compare-and-swap. |
| Pause/resume | **Authority / proof:** signed-in controls call `PATCH /v1/iris/adapter/pause?role=settings`; durable Iris persists the opt-out. Tests prove continuity `off`, `409 iris_memory_paused`, two contexts, resume, and a later write. **Limit:** the second context proves read-back after reload, not real-time push to open devices. |
| Single-fact delete | **Authority / proof:** a synced row sends settings role, hard mode, reason, and `confirmed=true`; durable Iris tombstones the id. Tests prove the real delete, local mirror, and next-review absence. **Limit:** a hard delete needs a password step-up recorded on the session in the last ten minutes (H13.b), but the route still defaults a missing `confirmed` to true. |
| Full clear | **Authority / proof:** `POST /v1/iris/adapter/memory/full-clear` needs the `FORGET EVERYTHING` confirmation, the forget role, and a fresh password step-up, and records a durable seven-day grace in Iris; `DELETE` on the same route cancels it. The page refuses the request when Iris is not reachable. **Limit:** the clear runs on the member's next Iris read or write after the grace ends, not on a timer, and no suite named here was rerun for this change. |
| Export | **Authority / proof:** `POST /v1/iris/adapter/export?role=settings` returns receipt metadata; the browser separately composes the bundle. Tests cover id/count/checksum plus the local artifact and audit chain. **Limit:** the receipt does not make the downloaded artifact a server-owned DSAR export. |

## Evidence map

Solid edges show the customer-facing sequence exercised across the named specs.
Dashed edges name the production authority the current surface does not yet
provide. Pause/resume, deletion and the full-clear request mutate durable Iris
authority; edit/conflict remains a local projection.

```mermaid
flowchart TB
    A[Signed-in profile] --> B[Hydrate Iris review, consent, and pause]
    B --> C[Browser control projection and local cache]
    C --> D[Edit and resolve same-origin conflict]
    D --> E[Pause or resume durable writes]
    E --> F[Delete one synced fact]
    F --> G[Request seven-day full clear]
    G --> H[Export receipt and browser bundle]
    D -. missing .-> I[Server edit and revision CAS]
    G -. not scheduled .-> K[Timed account-wide clear job]
    H -. not delivered .-> L[Server-owned export artifact]
```

## Proven observations

### Hydration clearly separates current Iris state from a cached projection

- `ProfileMemoryControls` starts from an honestly empty production state, then
  requests Iris review, consent, and pause data. A successful non-empty review
  maps server records into the control projection; the component does not ship
  fabricated customer memories as initial facts.
- The browser persists its projection under the exact key
  `oshun-profile-memory-controls-v1`. When review is unavailable, retained
  cached rows remain readable beside an explicit unavailable state and retry.
  When Iris returns no records, the rendered outcome still depends on the local
  projection and pause response rather than proving that every cached row was
  authoritatively erased.
- Category and scope chips, fact detail, provenance, recall history, downstream
  counts, and the local audit-chain link operate on the hydrated browser model.
  They are useful inspection affordances, but selecting a row does not issue a
  second server read.

### Edit and conflict resolution are deliberately browser-local

- `handleEditSubmit` compares the editor's expected revision against browser
  state, calls the local edit reducer, writes the next projection to
  localStorage, and emits redacted browser telemetry. It does not send a BFF
  edit request even when the selected row originated in Iris.
- The edit reducer supersedes the displayed item and creates a child revision
  with `editedFromItemId`. The resulting audit chain describes operations in
  this browser model; it is not a durable server ledger for the Iris record.
- The component listens for the same storage key's `storage` event. Two tabs on
  one origin can therefore expose a stale-draft conflict and choose the remote
  localStorage revision or layer another local revision on it. Two devices do
  not share that CAS boundary.

### Pause and single-record deletion reach durable Iris authority

- Signed-in pause toggles call the Iris adapter and commit the returned state
  only after success. The offline case fails closed: the UI does not pretend a
  pause was accepted when the mutation failed.
- The real-BFF suite proves the behavioral gate, not just the toggle. Continuity
  reports memory `off`; a live remember attempt returns
  `409 iris_memory_paused`; the blocked sentence is absent from review; resume
  persists `paused: false`; and a later remember call creates a visible row.
- Two isolated browser contexts with the same BFF identity read the same
  account-scoped pause after navigation/reload. That is durable read-back, not
  a claim of socket-driven synchronization between already-mounted pages.
- Deleting a synced fact sends hard mode, a reason, and `confirmed=true` to the
  management route. The durable Iris adapter tombstones the id, and a later
  review hides it. Only after a successful response does the component mirror
  the deletion into its local reducer.

### Full clear and export combine unlike authorities in one surface

- Since H13.a and H13.b (2026-09-24), typing `FORGET EVERYTHING` and
  confirming with the member's password first records a step-up on the BFF
  session, then posts the request to the Iris full-clear route, which stores a
  durable seven-day grace. Cancelling calls `DELETE` on the same route. When
  Iris is not reachable the page refuses the request; only cancelling keeps a
  browser fallback. The durable store applies a due clear the next time the
  member's memory is read or written, removing working and archival memories
  and keeping consents and opt-outs.
- For synced state, export first posts to the Iris adapter and displays a
  non-secret receipt with id, count, and checksum metadata. The downloadable
  “signed” bundle is then assembled from the browser control projection by
  `exportIrisUserMemoryBundle`; the server does not return that artifact's
  bytes through this flow.
- `profile-memory-controls.spec.ts` contains four local/stub-driven cases.
  `memory-edit-pause-forget.spec.ts` contains sixteen cases spanning anonymous
  gating, failure/retry, local conflicts, real Iris hydrate/export/delete,
  pause/read-back, two contexts, and the blocked-write boundary.

## Boundaries and gaps

- **Cached rows are not a fresh Iris read.** Unavailable and retry states must
  stay visible whenever the page is showing retained browser data.
- **A local revision is not a server edit.** The UI cannot promise that its
  revised content changed the durable Iris fact.
- **A same-origin conflict is not cross-device compare-and-swap.** The storage
  event model coordinates tabs sharing one browser origin, not independent
  devices or BFF writers.
- **A password step-up is a session record, not a per-action credential.** The
  BFF re-checks the password and records the time on the session; any
  destructive memory route accepts it for ten minutes, and `confirmed` still
  defaults to true when absent at the server boundary.
- **A due full clear runs on the next Iris access, not on a timer.** The grace
  record is durable, and reads after the deadline no longer show cleared
  memories, but no background job runs the clear at the deadline.
- **A tombstone mirror is not a second deletion.** The server tombstone is
  authoritative for a synced id; the local reducer updates the projection only
  after that mutation succeeds.
- **An export receipt is not the downloaded bundle.** Server metadata and the
  browser-composed artifact have distinct provenance and assurance levels.
- **A memory export is not a DSAR lifecycle.** The separate data-rights journey
  owns request, status, cancellation, and delivery claims.
- **Historical `/v1/memory` smoke is not current Iris proof.** The present flow
  uses `/v1/iris/adapter/*`; the 2026-05-29 walker cannot substantiate later
  route behavior.
- **Source reconciliation is not execution.** Neither cited application suite
  was rerun for this documentation audit or for the 2026-09-24 reconciliation.

## Re-run evidence

Run the two focused suites sequentially against one production-like web/BFF
pair and one immutable commit:

```bash
pnpm --dir apps/oshun/web exec playwright test -c playwright.config.ts e2e/profile-memory-controls.spec.ts --project=chromium --workers=1
pnpm --dir apps/oshun/web exec playwright test -c playwright.config.ts e2e/memory-edit-pause-forget.spec.ts --project=chromium --workers=1
```

Retain the commit, browser and BFF reports, backing-store tier, seeded identity,
and before/after Iris snapshots. A stronger follow-up should also retain a
restart result, two real-device traces, the password step-up exchange
(`POST /v1/auth/step-up`, and `403 step_up_required` without it), and a durable
full-clear request through its execution after the grace window.

## Source trail

- [Current journey contract](../journeys/memory-edit-pause-forget.md)
- [Current coverage registry](../journeys/coverage.md)
- [Profile memory controls](../../apps/oshun/web/src/app/profile/memory/ProfileMemoryControls.tsx)
- [Profile memory state bridge](../../apps/oshun/web/src/app/profile/memory/memory-state.ts)
- [Browser control reducers](../../libs/oshun/memory-iris/src/user-memory-controls.ts)
- [Iris adapter route](../../apps/oshun/bff/src/routes/iris.ts)
- [Iris adapter and durable store](../../apps/oshun/bff/src/adapters/iris-memory-adapters.ts)
- [Durable Iris startup wiring](../../apps/oshun/bff/src/server.ts)
- [Local control spec](../../apps/oshun/web/e2e/profile-memory-controls.spec.ts)
- [Real-BFF memory spec](../../apps/oshun/web/e2e/memory-edit-pause-forget.spec.ts)

## Cross-references

- [Profile memory view evidence](../customer/09-account/profile-memory.md)
- [Account deletion and DSAR journey](../journeys/account-deletion-and-dsar.md)
- [Account deletion and DSAR result](./account-deletion-and-dsar.md)
- [Assistant handoff result](./assistant-handoff-context-carry.md)
- [Crisis-aware tone-policy result](./crisis-aware-tone-policy.md)
- [Current Docs Center coverage audit](../../DOCS_CENTER_PAGE_AUDIT_2026-07-16.md)

## Open questions

- Which server edit contract will accept revision preconditions and return a
  durable conflict that independent devices can resolve safely?
- Should a destructive memory change bind its step-up to that one action rather
  than to the session's last ten minutes?
- Should a due full clear run on a timer and give the member a completion
  receipt, rather than applying on their next Iris access?
- Should an authoritative empty Iris review clear cached rows, quarantine them,
  or preserve them under an explicitly historical/offline label?
- Will export return a server-owned artifact with durable retention and access
  controls, or remain separate from the account DSAR workflow?
- When will “forget all mentions” across authorized actors gain a real UI,
  policy contract, and multi-record server operation?
