tagmem 캠페인 통합 — 무엇이 성능을 막고 있나

1 캠페인 연대기 — 무엇을 시도했고 무엇이 남았나

2 남은 오답 전수 해부 (flash 최고 구성 r4, 6/18행) — 직접 보고 판단

카드마다: 질의·GT·tagmem/vis 답 → 실제 검색된 entry(이미지 실물) → 타임라인 → reasoning 전문 → 원인 주석. vis ✓/✗ 병기.

find:val_144__2:q02 — 거짓 저장 (쓰기 환각) · vis ✓

질의: "Pick up the cell phone" GT: cell phone @ bedroom/dresser tagmem 답: cell phone @ living room/desk | vis 답: bedroom/dresser 검색: ['cell phone'] → 4개 entry

TIMELINE 'cell phone': [0s navigate@bedroom/dresser] -> [10s pick@bedroom/dresser] -> [83s navigate@bedroom/bed] -> [208s navigate@living room/desk]

[0s img] Robot approached dresser with table lamp, bowl, and cell pho
obj=table lamp,bowl,cell phone act=navigate room=bedroom surf=dresser

[10s] Robot picked up the cell phone from the dresser.
obj=cell phone act=pick room=bedroom surf=dresser

[83s img] Robot approached the bed with a remote control and pillows o
obj=remote control,pillow,cell phone act=navigate room=bedroom surf=bed

[208s img] Robot in living room with desk, desktop computer, cell phone
obj=desktopcomputer,cell phone,CD,house plant,remote control act=navigate room=living room surf=desk

tagmem reasoning 전문

The cell phone was moved from the dresser in the bedroom and was last seen resting on the desk in the living room at t=208s.

그래프의 "[10s] picked up the cell phone"은 거짓이다 — 실제로 집은 건 bowl이고 cell phone은 서랍장에 그대로였다. reader가 이 거짓 사건을 따라 '거실로 가져갔다'고 추론. 0s 스냅샷 이미지엔 진실(서랍장 위 cell phone)이 찍혀 있는데 글이 이겼다. 쓰기 규칙(근거쌍·인스턴스)으로 고치려 했으나 재-ingest 노이즈 대역(±0.06)에 묻힘. vis는 글이 없어서 픽셀만 보고 정답.

find:val_144__2:q04 — 저장 미스 + 지각 공통 · vis ✗

질의: "Pick up the pot" GT: pot @ kitchen/countertop tagmem 답: (기권) | vis 답: — 검색: ['pot'] → 16개 entry · FALLBACK

[0s img] Robot approached dresser with table lamp, bowl, and cell pho
obj=table lamp,bowl,cell phone act=navigate room=bedroom surf=dresser

[70s img] Bowl placed in the armchair next to the bed.
obj=bowl act=place room=bedroom surf=armchair

[83s img] Robot approached the bed with a remote control and pillows o
obj=remote control,pillow,cell phone act=navigate room=bedroom surf=bed

[104s img] Robot picking up the remote control from the bed.
obj=remote control act=pick room=bedroom surf=bed

[116s img] Robot entering bathroom near toilet and plunger.
obj=toilet,toilet tissue,plunger act=navigate room=bathroom surf=floor

[116s img] Robot approaching toilet in bathroom.
obj=toilet,toilet tissue,bar soap,plunger act=navigate room=bathroom surf=toilet

[175s img] Robot placed the remote control on the toilet lid.
obj=remote control,toilet,toilet tissue,bar soap,sponge act=place room=bathroom surf=toilet

[196s img] Robot picking up the remote control from the toilet lid.
obj=remote control,toilet act=pick room=bathroom surf=toilet

[208s img] Robot in living room with desk, desktop computer, cell phone
obj=desktopcomputer,cell phone,CD,house plant,remote control act=navigate room=living room surf=desk

[339s img] Robot placed the remote control on the round table next to t
obj=remote control,CD act=place room=living room surf=table

[355s] Robot entered the kitchen.
act=navigate room=kitchen

[355s img] Robot arrived at kitchen countertop with apple, soap dispens
obj=apple,soapdispenser,bowl,sponge,atomizer,toaster,bottle act=navigate room=kitchen surf=countertop

[406s] Robot picked up the bowl from the kitchen countertop.
obj=bowl act=pick room=kitchen surf=countertop

[406s img] Robot entered the living room holding a bowl.
obj=bowl,house plant act=navigate room=living room surf=floor

[495s] Robot placed the bowl on the table.
obj=bowl act=place room=living room surf=table

[495s img] Robot placed the bowl on the small round table near the desk
obj=bowl,desktopcomputer act=place room=living room surf=table

tagmem reasoning 전문

No pot was observed or logged in any room during the episode.

주방 스냅샷(355s)에 pot이 보이지만 태그 목록(apple, soapdispenser, bowl, sponge, atomizer, toaster)에서 pot만 누락 — R3 폴백으로 이미지를 전부 보여줘도 오답. vis도 오답 — 넓은 샷의 작은 pot은 지각 공통 한계.

find:val_406__0:q00 — twin 사슬 분리 실패 · vis ✓

질의: "Pick up the fork I moved earlier" GT: fork @ kitchen/dining table tagmem 답: fork @ living room/tv stand | vis 답: kitchen/dining table 검색: ['fork'] → 13개 entry

TIMELINE 'fork': [11s pick@kitchen/countertop] -> [23s navigate@living room] -> [71s place@living room/dresser] -> [155s look@dresser] -> [167s pick@dresser] -> [167s pick@dresser] -> [179s place@kitchen/dining table] -> [179s place@kitchen/dining table] -> [247s pick@kitchen/dining table] -> [247s pick@kitchen/dining table] -> [263s place@living room/tv stand] -> [263s place@living room/tv stand] -> [294s pick@living room/tv stand]

[11s] Picked up a fork from the countertop.
obj=fork act=pick room=kitchen surf=countertop

[23s] Entered the living room while holding the fork.
obj=fork act=navigate room=living room

[71s] Placed the fork on the dresser.
obj=fork act=place room=living room surf=dresser

[155s img] View of dresser with light blue container, fork, and books
obj=fork,vase,atomizer,candle,bottle,cup,newspaper act=look surf=dresser

[167s] Picked up the fork from the dresser.
obj=fork act=pick surf=dresser

[167s img] Robot picking up the fork from the dresser.
obj=fork act=pick surf=dresser

[179s] Placed the fork on the dining table.
obj=fork act=place room=kitchen surf=dining table

[179s img] Fork placed on the dining table alongside wine bottle, boile
obj=fork,wine bottle,boiler,atomizer,bread act=place room=kitchen surf=dining table

[247s] Picked up the fork from the dining table.
obj=fork act=pick room=kitchen surf=dining table

[247s img] Robot picking up a fork from the dining table.
obj=fork act=pick room=kitchen surf=dining table

[263s] Placed the fork on the tissue box on the tv stand.
obj=fork,tissue box act=place room=living room surf=tv stand

[263s img] Robot placing the fork on a tissue box on the tv stand.
obj=fork,tissue box act=place room=living room surf=tv stand

[294s img] Robot picking up the tissue box from the tv stand.
obj=tissue box,fork act=pick room=living room surf=tv stand

tagmem reasoning 전문

The robot moved the fork through several locations (countertop -> dresser -> dining table) before placing it on top of the tissue box on the tv stand in the living room at t=263s.

fork가 twin(옆에 같은 fork)인 에피소드. 검색은 13개 entry로 사슬 전체를 가져왔지만 타임라인이 두 fork의 사슬을 하나로 합쳐 마지막 hop을 엉뚱한 데(tv stand)로 짚음. 인스턴스 구분이 태그에 없어서(fork #1/#2 미표기) 결정론 분리 불가. vis는 정답.

find:val_733__0:q01 — 인스턴스 미저장 · vis ✓

질의: "Pick up all the remote controls I moved earlier" GT: remote control @ bedroom/dining table
remote control @ bathroom/toilet tagmem 답: remote control @ bedroom/dining table | vis 답: bedroom/dining table
bathroom/toilet 검색: ['remote control'] → 5개 entry

TIMELINE 'remote control': [97s navigate@living room/tv stand] -> [127s pick@living room/tv stand] -> [225s place@bedroom/dining table] -> [225s place@bedroom/dining table] -> [276s navigate@bedroom/dining table]

[97s img] Red vase and remote control on the tv stand in the living ro
obj=vase,remote control act=navigate room=living room surf=tv stand

[127s] The robot picked up a remote control from the tv stand in the living room.
obj=remote control act=pick room=living room surf=tv stand

[225s] The robot placed the remote control on the dining table in the bedroom.
obj=remote control act=place room=bedroom surf=dining table

[225s img] Remote control placed on the dining table in the bedroom.
obj=remote control,house plant act=place room=bedroom surf=dining table

[276s img] Dining table in the bedroom with a house plant, mug, remote
obj=dining table,house plant,mug,remote control,bed,laptop act=navigate room=bedroom surf=dining table

tagmem reasoning 전문

The timeline shows one remote control was moved. It was picked up from the tv stand in the living room and placed on the dining table in the bedroom, where it remained.

remote control 2개 중 #2의 이동(→toilet)이 그래프에 아예 없다 — 쓰기 모델이 '이미 저장한 것과 중복'으로 눌러버림. 검색이 아무리 좋아도 없는 건 못 가져온다. w3 쓰기(인스턴스 규칙)는 이 행을 고치지만 다른 행을 깨서 순증 0 (n=18 노이즈). vis는 정답.

find:val_733__0:q03 — 소물체 지각 (distractor) · vis ✓

질의: "Pick up the spoon" GT: spoon @ kitchen/countertop tagmem 답: spoon @ kitchen/dining table | vis 답: kitchen/countertop 검색: ['spoon'] → 3개 entry

TIMELINE 'spoon': [0s pick@kitchen/countertop] -> [9s place@kitchen/countertop] -> [411s navigate@kitchen/dining table]

[0s img] Robot picks up a bowl from the countertop near a soap dispen
obj=bowl,soapdispenser,apple,spoon act=pick room=kitchen surf=countertop

[9s img] A bowl placed on the kitchen countertop next to an apple and
obj=bowl,soapdispenser,apple,spoon act=place room=kitchen surf=countertop

[411s img] Robot approached the dining table with a cell phone, house p
obj=dining table,cell phone,house plant,bottle,vase,fork,knife,spoon,chair act=navigate room=kitchen surf=dining table

tagmem reasoning 전문

The timeline and memory entries indicate that a spoon is located on the dining table in the kitchen.

spoon은 광폭 스냅샷에 태그돼 검색은 성공(0s·9s countertop 이미지에 spoon이 실제로 찍혀 있음). reader가 나중 dining table 이미지(포크·나이프 있는)로 끌려가 오답. 작은 물체 + 유사 맥락 혼동 — vis는 16장 중에서 맞는 프레임을 찾아냄.

find:val_733__0:q06 — 지각 공통 (vis도 오답) · vis ✗

질의: "Pick up the vase" GT: vase @ living room/tv stand tagmem 답: vase @ kitchen/chair | vis 답: kitchen/dining table 검색: ['vase'] → 8개 entry

TIMELINE 'vase': [97s navigate@living room/tv stand] -> [110s pick@living room/tv stand] -> [241s navigate@living room/cart] -> [411s navigate@kitchen/dining table] -> [475s pick@kitchen/dining table] -> [475s pick@kitchen/dining table] -> [520s place@kitchen/chair] -> [520s place@kitchen/chair]

[97s img] Red vase and remote control on the tv stand in the living ro
obj=vase,remote control act=navigate room=living room surf=tv stand

[110s] The robot picked up a vase from the tv stand in the living room.
obj=vase act=pick room=living room surf=tv stand

[241s img] Cart in the living room with a bowl near it and a red vase o
obj=cart,vase,bowl act=navigate room=living room surf=cart

[411s img] Robot approached the dining table with a cell phone, house p
obj=dining table,cell phone,house plant,bottle,vase,fork,knife,spoon,chair act=navigate room=kitchen surf=dining table

[475s] The robot picked up a vase from the dining table.
obj=vase act=pick room=kitchen surf=dining table

[475s img] The robot picking up a white vase or bottle from the dining
obj=vase,bottle act=pick room=kitchen surf=dining table

[520s] The robot placed a vase on the chair.
obj=vase,bottle act=place room=kitchen surf=chair

[520s img] The robot placed a white vase on the green chair.
obj=vase,bottle,chair act=place room=kitchen surf=chair

tagmem reasoning 전문

The white vase was picked up from the dining table in the kitchen and placed on the green chair at t=520s, where it remains in the final view.

97s 이미지에 vase@tv stand(=GT)가 찍혀 있고 검색도 됐다. 그런데 reader가 다른 위치로 답함 — vis도 같은 행에서 오답. 에피소드 내 vase 유사물/시점 모호성의 지각 공통 한계.

3 대조: 회복에 성공한 카드 3장

find:val_144__2:q01 — 읽기 v2가 회복 (텍스트 앵커링 해제) · vis ✗

질의: "Pick up the remote control I moved earlier" GT: remote control @ living room/dining table tagmem 답: remote control @ living room/dining table | vis 답: living room/cart 검색: ['remote control'] → 11개 entry

TIMELINE 'remote control': [83s navigate@bedroom/bed] -> [104s pick@bedroom/bed] -> [104s pick@bedroom/bed] -> [175s place@bathroom/toilet] -> [175s place@bathroom/toilet] -> [196s pick@bathroom/toilet] -> [196s pick@bathroom/toilet] -> [208s navigate@living room] -> [208s navigate@living room/desk] -> [339s place@living room/table] -> [339s place@living room/table]

[83s img] Robot approached the bed with a remote control and pillows o
obj=remote control,pillow,cell phone act=navigate room=bedroom surf=bed

[104s] Robot picked up the remote control from the bed.
obj=remote control act=pick room=bedroom surf=bed

[104s img] Robot picking up the remote control from the bed.
obj=remote control act=pick room=bedroom surf=bed

[175s] Robot placed the remote control on the toilet lid.
obj=remote control,toilet act=place room=bathroom surf=toilet

[175s img] Robot placed the remote control on the toilet lid.
obj=remote control,toilet,toilet tissue,bar soap,sponge act=place room=bathroom surf=toilet

[196s] Robot picked up the remote control from the toilet lid.
obj=remote control,toilet act=pick room=bathroom surf=toilet

[196s img] Robot picking up the remote control from the toilet lid.
obj=remote control,toilet act=pick room=bathroom surf=toilet

[208s] Robot entered the living room carrying the remote control.
obj=remote control act=navigate room=living room

[208s img] Robot in living room with desk, desktop computer, cell phone
obj=desktopcomputer,cell phone,CD,house plant,remote control act=navigate room=living room surf=desk

[339s] Robot placed the remote control on the table.
obj=remote control act=place room=living room surf=table

[339s img] Robot placed the remote control on the round table next to t
obj=remote control,CD act=place room=living room surf=table

tagmem reasoning 전문

The robot picked up the remote control from the bed in the bedroom, temporarily placed and picked it back up from the toilet lid in the bathroom, and finally placed it on the round table in the living room. The allowed surface corresponding to this table is 'dining table'.

remote control — entry 원문이 "round table"이라 계속 desk로 답하던 행. '글보다 픽셀' 규칙 + LATEST 주석 후 이미지에서 dining table을 읽어냄. vis는 이 행 오답.

find:val_733__0:q05 — 읽기 v4가 회복 (타임라인 세기) · vis ✗

질의: "Pick up all the forks" GT: fork @ kitchen/dining table
fork @ kitchen/dining table tagmem 답: fork @ kitchen/dining table
fork @ kitchen/dining table | vis 답: kitchen/dining table 검색: ['fork'] → 17개 entry

TIMELINE 'fork': [411s navigate@kitchen/dining table]

[0s] The robot picked up a bowl from the kitchen countertop.
obj=bowl act=pick room=kitchen surf=countertop

[0s img] Robot picks up a bowl from the countertop near a soap dispen
obj=bowl,soapdispenser,apple,spoon act=pick room=kitchen surf=countertop

[9s] The robot placed a bowl on the kitchen countertop.
obj=bowl act=place room=kitchen surf=countertop

[9s] The robot picked up a bowl from the kitchen countertop.
obj=bowl act=pick room=kitchen surf=countertop

[9s img] A bowl placed on the kitchen countertop next to an apple and
obj=bowl,soapdispenser,apple,spoon act=place room=kitchen surf=countertop

[139s] The robot navigated to the dining table in the bedroom.
act=navigate room=bedroom surf=dining table

[139s img] House plant on the dining table with a bed in the bedroom.
obj=house plant,bed,chair act=navigate room=bedroom surf=dining table

[225s] The robot placed the remote control on the dining table in the bedroom.
obj=remote control act=place room=bedroom surf=dining table

[225s img] Remote control placed on the dining table in the bedroom.
obj=remote control,house plant act=place room=bedroom surf=dining table

[276s] The robot navigated to the dining table in the bedroom.
obj=dining table act=navigate room=bedroom surf=dining table

[276s img] Dining table in the bedroom with a house plant, mug, remote
obj=dining table,house plant,mug,remote control,bed,laptop act=navigate room=bedroom surf=dining table

[411s] The robot navigated to the dining table in the kitchen.
obj=dining table act=navigate room=kitchen

[411s img] Robot approached the dining table with a cell phone, house p
obj=dining table,cell phone,house plant,bottle,vase,fork,knife,spoon,chair act=navigate room=kitchen surf=dining table

[475s] The robot picked up a vase from the dining table.
obj=vase act=pick room=kitchen surf=dining table

[475s img] The robot picking up a white vase or bottle from the dining
obj=vase,bottle act=pick room=kitchen surf=dining table

[520s] The robot placed a vase on the chair.
obj=vase,bottle act=place room=kitchen surf=chair

[520s img] The robot placed a white vase on the green chair.
obj=vase,bottle,chair act=place room=kitchen surf=chair

tagmem reasoning 전문

There are two forks visible on the dining table in the kitchen (one on the left side near the place setting and one on the right side near the house plant). Neither fork was moved during the episode.

fork 2개 multi — TIMELINE이 두 사슬을 분리해 개수 2를 맞춤. vis는 오답 (같은 표면 2개를 프레임에서 못 셈).

find:val_144__2:q05 — R2 모델 태거가 회복 (동의어) · vis ✓

질의: "Pick up the spray bottle" GT: atomizer @ kitchen/countertop tagmem 답: atomizer @ kitchen/countertop | vis 답: kitchen/countertop 검색: ['bottle'] → 3개 entry

TIMELINE 'bottle': [355s navigate@kitchen/countertop]

[355s] Robot entered the kitchen.
act=navigate room=kitchen

[355s img] Robot arrived at kitchen countertop with apple, soap dispens
obj=apple,soapdispenser,bowl,sponge,atomizer,toaster,bottle act=navigate room=kitchen surf=countertop

[406s] Robot picked up the bowl from the kitchen countertop.
obj=bowl act=pick room=kitchen surf=countertop

tagmem reasoning 전문

The spray bottle (atomizer) was seen on the kitchen countertop at t=355s and remained there throughout the episode.

spray bottle → 그래프 태그는 atomizer. 규칙 태거 0히트 → 모델 태거 1콜이 vocab으로 번역해 히트.

4 v0.7 길이층에서 vis가 이긴 대표 사례

길이(초)를 4.5배 늘려도 vis가 안 깎인 이유가 카드에서 보인다 — 사건이 ~4-6개뿐이라 큐레이션 16장이 배치를 전부 덮는다. tagmem 역전 조건 = 사건 밀도 > 16.

v0.7 short · val_476__0 pepper shaker — vis ✓ / tagmem ✗

질의: "Pick up the pepper shaker" GT: pepper shaker @ kitchen/dining table tagmem 답: living room/dining table 그래프 11 entries 중 8개 검색

[0s img] Table with house plant, vase, bowl, plate, cell phone, pen,
obj=house plant,vase,bowl,plate,cell phone,pen,box,salt shaker act=look surf=dining table

[14s] Robot picked up salt shaker from dining table.
obj=salt shaker act=pick surf=dining table

[26s] Robot placed salt shaker on desk.
obj=salt shaker act=place surf=desk

[26s img] Desk with salt shaker, cell phone, and house plant.
obj=salt shaker,cell phone,house plant act=place surf=desk

[77s] Robot picked up salt shaker from desk.
obj=salt shaker act=pick surf=desk

[93s] Robot placed salt shaker on dining table.
obj=salt shaker act=place surf=dining table

[93s img] Dining table with salt shaker, bottle, house plant, bread, a
obj=salt shaker,bottle,house plant,bread,plate act=place surf=dining table

[112s img] Robot grasping vase on dining table.
obj=vase,bottle,house plant,salt shaker,bread,plate act=pick surf=dining table

reasoning 전문

The memory entries for [pepper shaker] show that the shaker (mistakenly labeled as salt shaker in captions) was initially on the dining table, moved to the desk, and then moved back to the dining table at t=93s where it remains. The dining table is located in the living room.

v0.7 long · val_772__0 candle — vis ✓ / tagmem ✗

질의: "Pick up the candle I moved earlier" GT: candle @ bedroom/dresser tagmem 답: living room/desk 그래프 36 entries 중 8개 검색

[213s] picked up green candle from dining table
obj=candle,atomizer act=pick room=living room surf=dining table

[213s img] robot picking up green cylinder from dining table
obj=candle,atomizer act=pick room=living room surf=dining table

[231s] placed green candle on small round side table
obj=candle,atomizer act=place room=living room surf=table

[231s img] green candle placed on small round side table
obj=candle,atomizer act=place room=living room surf=table

[243s] picked up green candle from round side table
obj=candle,atomizer act=pick room=living room surf=table

[243s img] robot carrying green candle over desk with cell phone and pe
obj=candle,atomizer,cell phone,pen act=navigate room=living room surf=desk

[318s] placed green candle on desk
obj=candle,atomizer act=place room=living room surf=desk

[318s img] green candle placed on desk next to cell phone
obj=candle,atomizer,cell phone act=place room=living room surf=desk

reasoning 전문

The memory log shows that at t=318s, the green candle was placed on the desk in the living room.

5 S3 대표본 검증 (353질의 × 2모델, $22.2) — 최종 리더보드 위치

커밋 d6f4499 구성(쓰기 v1 + 읽기 v2 + r3)으로 S3 표본 전체 측정. 방/표면 병기:

gemini-3.6-flash

claude-sonnet-5 (전체만)

vis 0.48/0.38 > frames 0.46/0.35 > textlog 0.45/0.41 > tagmem 0.41/0.33 > rf-v2 0.35/0.29 > blind 0.20/0.15

6 저장 제약 맞대결 — cap 16 → 8에서 순위 역전 (flash S3 353질의, vis 실측 재-ingest $15.77)

7 결론 — 성능을 막는 것 (수치로)