Ablation over the Memory Modules
w/o MemoryOS
w/o MTM
w/o LPM
w/o Chain
MemoryOS
LoCoMo · F1 by question type
10
20
30
40
50
15.23
35.27
Single Hop
12.12
41.15
Multi Hop
6.31
20.02
Temporal
17.39
48.62
Open Domain
GVD · three judged metrics
20
40
60
80
100
40.4
93.3
Acc.
46.6
91.2
Corr.
63.2
92.3
Cohe.
5
25.13
10
35.27
20
37.82
30
38.32
40
37.98
retrieved pages k
k
k = 10 adopted · gains flatten beyond 20
single-hop F1 on LoCoMo · 5 → 10 is the big step, 20 → 40 is flat or down (noise from extra pages)
footnote · on temporal questions the chain is the one exception: 20.34 without it vs 20.02 with it · all values GPT-4o-mini, Fig. 2 and Fig. 3
MTM
matters most ·
LPM
next ·
chain
least · every tier removed costs something on every question type
24 / 25