Whether a stored memory helps answer a user's question often depends on world knowledge rather than surface similarity. Jetriever replaces similarity search with full-pool utility scoring: given a query, it uses Jev, an LLM-derived System One decision model, to score every stored memory for the probability that it is useful to the answer, ranking the memory pool directly by those probabilities. On InMind's 125 indirect-query tasks, Jetriever places the gold exchange first in 84.8% of cases (versus 1.6% for BM25 and 0.0% for MiniLM with user-turn indexing) and reaches 76.0% end-to-end Application, compared with at most 16.0% across published retrieval and memory systems. A controlled 20-question comparison disentangles recognizing implicit evidence from ranking it properly: Jev and Qwen3.5-9B both recover 90% of gold exchanges within the top 50 (versus 30% for the BERT-based decision model Laya), yet at rank 1 Jev reaches 85% while Qwen3.5-9B drops to 40%. On a dimension-proportional random sample of 60 MemoryAgentBench questions, Jetriever raises the four-dimension average from 44.3% (ReFind) to 55.4%, at a mean retrieval cost of USD 0.00834 per InMind query and USD 0.03103 per MAB query. Together, these results show that effective utility retrieval hinges on combining broad world knowledge, calibrated ranking probabilities, and affordable full-pool evaluation.
Figures & tables
Method and indexed text
R@1
R@5
R@10
BM25, user turn
1.6
4.8
5.6
MiniLM, user turn
0.0
2.4
8.8
BM25, full exchange
0.8
2.4
4.0
MiniLM, full exchange
0.0
1.6
7.2
Jetriever, full exchange
84.8
92.0
94.4
Table 1: InMind-125 gold-exchange recall (%). All methods rank the identical pool of 241 candidate exchanges per query; user-turn controls index only user utterances, whereas Jetriever evaluates full user–assistant exchanges.
Configuration
Application (%)
Backbone control
GPT-5-mini, target in context
84.0
Naive RAG
BM25
9.6
text-embedding-3-large
16.0
Memory systems
Table 2: InMind-125 end-to-end Application scores (%). Published baselines are taken from InMind [ 1 ] ; Jetriever is evaluated on the same 125 tasks with the same GPT-5-mini backbone and context-aware Application criterion. See Appendix A.1 for the complete table.
Method
AR
TTL
LRU
SF
Overall
GPT-4o-mini
49.2
48.6
46.2
25.0
42.2
BM25
60.5
44.5
35.6
25.5
41.5
text-embedding-3-large
54.6
44.3
37.3
16.0
38.0
HippoRAG 2 [ 26 ]
65.1
35.8
36.2
29.5
41.6
MIRIX [ 28 ]
47.5
24.1
25.4
8.0
26.2
ReFind ∗ [ 3 ]
49.8
30.0
55.5
41.7
44.3
Table 3: MAB competency scores (%), all using GPT-4o-mini for answer generation. Unstarred rows are published full-benchmark baselines from [ 2 ] . ∗ ReFind and Jetriever are evaluated head-to-head on the same dimension-proportional 60-question random subset (within AR, proportional sampling drew EventQA, MH-QA, and LME items, with no SH-QA item; dimension means aggregate the drawn subtasks).
Figure 1: Gold-exchange R@k at k∈{1,2,5,10,20,50} across the same 20 InMind tasks (241 candidate exchanges per task; logarithmic horizontal axis). Markers reflect exact per-task gold ranks; Qwen3.5-4B [ 15 ] is included under the same scoring protocol as Qwen3.5-9B .
Run
Queries
Mean candidates
USD/query
InMind exchanges
125
241
0.00834
InMind sessions
125
47
0.00574
MAB random conversations
60
1,258.2
0.03103
Table 4: Observed Jev API retrieval charges (excluding downstream answer generation). Total retrieval spend is 1.0425forInMindexchanges,0.7179 for InMind sessions, and $1.8617 across 75,532 scored segments on MAB-60.
Jev (System One model)
Generative LLM scorer
BERT-based classifier
Broad world knowledge
✓
✓
×
Calibrated ranking scores
✓
×
✓
Full-pool cost and latency
Low–moderate
High
Low
Table 5: Comparison of scorer families across the three requirements for full-pool utility retrieval.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Configuration
Application (%)
Backbone control
GPT-5-mini, target in context
84.0
Naive RAG
MiniLM
9.6
text-embedding-3-large
16.0
BM25
9.6
Appendix
Table 6: All 16 published InMind Application configurations [ 1 ] , followed by Jetriever’s top-five result. All rows use the same context-aware Application scoring criterion.
Original MAB baseline
AR
TTL
LRU
SF
Overall
Long-context agents
GPT-4o (128K)
58.1
50.0
54.9
32.5
48.8
GPT-4o-mini (128K)
49.2
48.6
46.2
25.0
42.2
Claude-3.7-Sonnet (200K)
59.7
53.9
62.2
22.5
49.6
GPT-5-mini (400K)
74.4
48.6
66.2
53.0
60.6
GPT-4.1-mini (1M)
71.8
46.2
49.1
20.5
46.9
Appendix
Table 7: All 23 baselines in the original MAB main table [ 2 ] : four competency-group averages and Overall (%). AR = Accurate Retrieval; TTL = Test-Time Learning; LRU = Long-Range Understanding; SF = Selective Forgetting. These are full-benchmark scores rather than MAB-60 subset scores.
Published method
SH-QA
MH-QA
LME
EventQA
FC-SH
FC-MH
Avg.
GPT-4o-mini
64.0
43.0
30.7
59.0
45.0
5.0
41.1
BM25-RAG
66.0
56.0
45.3
74.6
48.0
3.0
48.8
Contriever
22.0
31.0
15.7
66.8
18.0
7.0
26.8
text-embed-3-small
60.0
44.0
48.3
63.0
28.0
3.0
41.1
text-embed-3-large
54.0
44.0
50.3
70.0
28.0
4.0
41.7
Qwen3-Embed-4B
57.0
47.0
43.3
71.4
29.0
3.0
41.8
Appendix
Table 8: Published full six-subtask comparison for Accurate Retrieval (SH-QA, MH-QA, LME, EventQA) and Selective Forgetting (FC-SH, FC-MH), reproduced from ReFind [ 3 ] , with reused baselines from MAB [ 2 ] . Avg. is the unweighted six-subtask mean (%). This table reports full task sets rather than our 60-question subset.