[Frontier] [MEM16-01A] Carry precise file ancestry through memory and later conversations #402

Parent

S16 #392, MEM16-01 delivery graph; source-lineage dependency of FILE16-01 #396.

Model tier

Frontier: transitive private ancestry across memory, native/history derivatives, recipient approval and revocation is a protected contract.

Context to read

AGENTS.md, CONTEXT.md, docs/agents/workflow.md, S16/S15 specs, ADRs 0064/0074/0076, docs/design/S16-rakazo-experience.md, docs/contracts/s16-files-v1.md, #400/#401. Audit src/s15/{memory,team,disclosure}/, conversation/{application,types,owner-view}.ts and src/s16/files/. Preserve current FR/NFR and source/forget semantics.

Deliverable

Useful file-derived memory can inform a later authorized conversation or Bot without losing the exact original/derivative file ancestry. Forget, correction and source withdrawal invalidate uncited descendants and future model/tool/recipient sends. Replace the explicit text-slice file-memory/history fence only when these checks are composed.

Implementation steps

1. Freeze metadata-only source/claim/packet ancestry queries with exact bounded identities; no serialized principal, generic resource label or model citation substitutes for a source.
2. Bind file metadata anchors and actual observed content claims to protected MEM/TEAM provenance, retention scope and current pinned resources. Derive summaries/artifacts from complete ancestry.
3. Track precise file ancestors of every included memory/history fragment before interpreter/model/tool/VM release and recipient selection. Unrelated withdrawn files must not block clean unrelated work.
4. Make file withdrawal/forget/correct monotonic across MEM, TEAM, CHAT result/history/native derivatives, artifacts, previews/downloads and replay suppression. Define conservative cross-ledger commit/recovery; evaluate a shared SQLite writer when it simplifies atomicity rather than adding ad-hoc distributed flags.
5. Allow future authorized requests/Bots to use file-derived memory only through current private/team scope and precise new-recipient approval.
6. Implement authored source/impact/correct/forget feedback with actual service facts and Thai/English keys; remove the unavailable-memory capability only after integration passes.

Acceptance criteria

• A later request uses an actual useful file-derived fact with source provenance, within current scope, without explicitly reattaching the entire file.
• Private data does not move to another Bot/resource/recipient via uncited memory, summary, interpreter history, native compaction, artifact or handoff.
• Withdrawal/forget during a live wait prevents the next send/publication; future retrieval/history excludes retired content and cannot replay it after restart.
• Correction preserves inherited ceilings and invalidates stale dependent outputs; replay suppression does not retain erased private bytes as a new memory source.
• Exact source approvals cannot be bypassed with a generic computer source; clean unrelated work remains usable.
• The first text-only contract remains compatible until a versioned enabled-memory capability is present.

Tests and evidence

1. Composed file → actual tool observation → MEM retention → later Bot/request retrieval → actual authorized model input.
2. Exact-source versus broad-resource recipient approval; private/team scope, withdrawn/corrected parent, uncited descendants and concurrent requests.
3. Forget/withdraw before and after retrieval/model wait/compaction/publication/download, restart and duplicate command identities.
4. Browser source/impact/correct/forget states use current protected projections, with no fabricated completion.

Requirements

FR-05/07/15/16/17/18/22/31/35/38/42; NFR-02/05/06/07/10/11/12/14; S16-AC03/04/08/10/20.

Blocked by

#400 precise file source/custody/publication. Coordinate original/derivative identities with #401; text-only memory proof can progress independently of a live parser/vision route. Contract/commit design must be reviewed before ready-for-agent.

Contracts used/delivered

Existing MEM/TEAM/DATA15/POLICY owning ports plus versioned metadata ancestry and file-dependent history/publication bindings. Schway public native/session/compaction surface only.

Out of scope

Global storage migration without preservation proof, unrelated connectors/native clients, new work authority, owner-host activation and private S14 data promotion.

Do not

Do not remove the current memory/history fence before transitive checks pass, trust model citations for ancestry, scan all old files as blanket lineage, store credentials, silently broaden recipient/resource scope, add model attribution, use implement-parallel or merge.

Definition of done

bun install --frozen-lockfile, bun run typecheck, relevant MEM/TEAM/CHAT/DATA15/file/composed/browser regressions and git diff --check pass. PR Closes this issue, Refs #392/#396, records exact contract/commit/evidence, and does not claim complete MEM16-01 search/UI/native coverage from source ancestry alone.

Complete backlog publication — 2026-11-05

Preserve the primary-owned prepared jit/402-s16-file-memory branch; no new implementation claim or edits are made here. #416 owns remaining memory search/client lifecycle; #407/#413/#421/#433 consume exact file ancestry. Consider a metadata query/commit contract before ad-hoc cross-ledger flags; do not remove the current memory/history fence before proof.