FME — Object Model (LOCKED)

Family Memory Engine. The product is "nothing gets forgotten." Source plan: .claude/plans/hazy-noodling-firefly.md. Locked 2026-06-05 (ZW-ENGINE-V9). This is law — every object in the engine conforms.

Single source of truth (no info in 5 places)

Original Artifact — the raw Telegram message / voice file / PDF / photo / link = the source of truth. Kept forever.
Family Memory Index — the searchable catalog (one row per object). Derived.
Family_Event_Log — immutable, append-only audit. Derived.

Everything else is derived from these three. Never duplicate authority.

Canonical object schema (every object, no exceptions)

id:           FLB-YYYYMMDD-NNNNN          # unique, sequential per day
type:         message | voice | photo | pdf | link | task | grocery | memory | event
created:      ISO-8601 timestamp
createdBy:    Chanie | Sam | kid | system
chapter:      ONE   (Kids | Travel | Home | Simchas | Grocery | Life-Expenses-Info | Unfiled)
entities:     MANY  [Libby, Camp Mesorah, Tuition]     # people & things — PRIMARY for retrieval
tags:         MANY  [Summer 2027, Registration]
visibility:   both | sam                   # family chapters = both; Sam-personal excluded from family store
source:       Telegram | portal | SMS | drive
original_uri: permanent ref to the RAW artifact — NEVER deleted
summary:      one-line (AI or raw)
content:      text / file-ref
status:       ACTIVE | ARCHIVED
origin:       Human | AI                    # who created it
review_state: AI_UNREVIEWED | CONFIRMED | CORRECTED
lifespan:     Permanent | Archive           # AI suggests Archive; humans decide

Rules (law)

Chapters AND entities AND tags — all three. One chapter, many entities, many tags. A camp PDF = chapter Kids + entities [Eli, Camp Mesorah] + tags [Summer 2027, Registration] → "show me Eli" / "Camp Mesorah" / "Kids" all resolve.
entities[] exists from DAY 1 (field present in Phase 1A even though AI auto-extraction is Phase 2 — never lose the field).
Unfiled is a first-class chapter, never an exception. Classify fails → Unfiled, acknowledge, continue.
Source preservation: keep original + transcript + summary (for voice, all three) — never collapse to only the summary.
Trust: every object records origin (Human/AI) + review_state so months later you know what the AI guessed vs what a human confirmed.

Where it lives — storage (CORRECTED 2026-06-05 for scale; Sam flagged Sheet row limits)

Live message threads = Cloudflare KV (chanie:thread etc., rolling 50-msg buffer). No row limit. Already live.
Family Memory Index (the catalog) = Cloudflare D1 (SQLite) — NOT a Google Sheet. Sheets cap ~10M cells / a few hundred-K rows and would run out; D1 holds millions of rows AND is SQL-queryable = the retrieval backbone. (A Sheet may mirror a recent slice for human browsing, but D1 is the system of record.)
Family_Event_Log (append-only, grows forever) = a Cloudflare D1 table — never truncated.
Originals = a dedicated Google Drive folder (every raw photo/PDF/voice forever); original_uri points at the Drive file.
🔴 PROTECTED: family store ONLY — never reaches Mildred's / any client's store (two-store wall).
Why D1: same Cloudflare stack as the workers · free tier generous · SQL = fast retrieval · no row ceiling. Provision via CF D1 tooling when Phase 1A's Family_Event_Log lands.