MemMend: A Selective Revision Framework for Memory Question Answering
Organizations: Independent Research
Abstract
Revising every answer from a long-term memory can consume substantial inference budget and damage facts that were already correct. We introduce MemMend, a modular selective revision framework that routes potentially incomplete answers to an editor, represents corrections as bounded exact-text replacements, and accepts only edits with resolvable source coordinates and valid edit operations. The editor preserves the current answer when no legal repair is proposed; accepted edits require no additional semantic-verification call. On 1,986 LoCoMo development questions, MemMend achieves 67.17 conversation-macro F1 and 88.35 mean Judge score, improving over a verified revision chain by 0.82 and 1.24 points while using 62.37% fewer revision calls. Routing uses 54.15% fewer revision calls than ungated local review, with similar F1 but a 0.67-point lower Judge score. Applied directly to native PropMem, Mem0, and SimpleMem answers, the editing interface raises F1 by 3.79, 6.44, and 0.42 points. It outperforms same-gate rewriting and saves 49.85-83.72% of editing calls relative to ungated editing, which retains higher F1 in this adaptation. A source-by-editing factorial supports local editing for F1 but leaves an independent quality benefit from source obligations unresolved. Transfer to 451 LongMemEval-V2 Small questions also favors revision efficiency, while an external PerLTQA dialogue subset reveals a completeness gap. These results support selective editing as a reusable quality-cost operating point.
Figures & tables
| Method | F1 | Acc. | Rel. | Comp. | Judge mean | Extra calls |
|---|---|---|---|---|---|---|
| Scope-checked answer | 63.96 | 81.37 | 91.49 | 77.74 | 83.53 | 0 |
| Premise Guard | 63.40 | 80.31 | 89.68 | 76.94 | 82.31 | 0 |
| Verified revision chain | 66.35 | 84.54 | 94.66 | 82.12 | 87.11 | 2,493 |
| Iterative revision | 48.22 | 75.38 | 87.36 | 70.39 | 77.71 | 5,958 |
| Critique and rewrite | 61.97 | 83.03 | 93.96 | 81.12 | 86.04 | 5,958 |
| MemMend | 67.17 | 85.40 | 96.12 | 83.53 | 88.35 | 938 |
| System | F1 | Judge |
|---|---|---|
| PropMem | 62.13 | 80.72 |
| Mem0 | 48.14 | 74.77 |
| SimpleMem | 46.95 | 76.85 |
| MemMend | 67.17 | 88.35 |
| F1 | Editing calls | ||||||
|---|---|---|---|---|---|---|---|
| Memory system | Native | SP | UP | SR | SP | UP | SR |
| PropMem | 62.13 | 65.92 | 66.47 | 64.88 | 916 | 1,998 | 910 |
| Mem0 | 48.14 | 54.58 | 56.58 | 53.72 | 997 | 1,988 | 996 |
| SimpleMem | 46.95 | 47.37 | 48.12 | 46.95 | 324 | 1,990 | 325 |
| Policy | Source | F1 | Judge |
|---|---|---|---|
| Scope-checked answer | – | 63.79 | 83.28 |
| Premise Guard | – | 62.99 | 81.67 |
| Local edit | Required | 66.86 | 88.03 |
| Local edit | Not required | 67.00 | 88.12 |
| Whole-answer rewrite | Required | 66.00 | 88.25 |
| Whole-answer rewrite | Not required | 66.22 | 88.42 |
| Method | Accuracy (%) | Correct | Extra calls | Tokens/question |
|---|---|---|---|---|
| Reader | 40.58 | 183/451 | 0 | 14,561 |
| MemMend (selective patching) | 41.24 | 186/451 | 57 | 16,359 |
| Ungated local review | 41.02 | 185/451 | 451 | 29,916 |
| Selective rewriting | 40.80 | 184/451 | 58 | 16,399 |
| Generic revision | 33.70 | 152/451 | 451 | 29,254 |
| Repeated generic revision | 32.82 | 148/451 | 902 | 43,946 |
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | F1 | Judge | Extra calls | F1 [95% CI] | Judge [95% CI] |
|---|---|---|---|---|---|
| MemMend | 67.17 | 88.35 | 938 | – | – |
| w/o restoration | 67.20 | 88.37 | 909 | ||
| Ungated local review | 67.09 | 89.02 | 2,046 | ||
| w/o edit bounds | 66.83 | 88.18 | 937 |
| Routing policy | F1 | Judge | Extra calls |
|---|---|---|---|
| Restoration only | 63.80 | 83.01 | 60.0 |
| MemMend routing | 67.17 | 88.35 | 938.0 |
| Ungated local review | 67.09 | 89.02 | 2046.0 |
| Random: matched calls | 65.21 | 85.59 | 938.0 |
| Random: matched tokens | 65.21 | 85.60 | 940.7 |
| Contrast | F1 [95% CI] | Judge [95% CI] |
|---|---|---|
| Source effect: local edit | ||
| Source effect: rewrite | ||
| Locality effect: source required | ||
| Locality effect: no source obligation | ||
| Average source effect | ||
| Average locality effect |
| Policy | Source | Changed | Adequate | Unsupported additions | Repairs | Harms |
|---|---|---|---|---|---|---|
| Local edit | Required | 16 | 55.97 | 1.23 | 2.02 | 0.00 |
| Local edit | Not required | 22 | 56.87 | 1.72 | 2.91 | 0.00 |
| Whole-answer rewrite | Required | 24 | 57.18 | 1.79 | 2.82 | 0.00 |
| Whole-answer rewrite | Not required | 20 | 56.86 | 1.72 | 2.50 | 0.00 |
| Comparator | Wins/losses | accuracy [95% CI] |
|---|---|---|
| Reader | 3/0 | |
| Ungated local review | 1/0 | |
| Selective rewriting | 2/0 | |
| Generic revision | 35/1 | |
| Repeated generic revision | 39/1 | |
| Critique and rewrite | 25/5 |
| Revision policy | Matched budget | Accuracy (%) | accuracy [95% CI] | |
|---|---|---|---|---|
| Generic revision | Calls | 0.1264 | 39.71 | |
| Generic revision | Tokens | 0.1223 | 39.74 | |
| Critique and rewrite | Calls | 0.0632 | 40.34 | |
| Critique and rewrite | Tokens | 0.0612 | 40.35 |
| Method | F1 | Judge | Extra calls | F1 [95% CI] | Judge [95% CI] |
|---|---|---|---|---|---|
| Scope-checked answer | 19.89 | 44.17 | 0 | ||
| Premise Guard | 18.54 | 41.50 | 0 | ||
| Verified revision chain | 20.07 | 44.67 | 286 | ||
| Iterative revision | 25.92 | 51.50 | 600 | ||
| Critique and rewrite | 21.21 | 45.50 | 600 | ||
| MemMend | 20.65 | 46.83 | 133 | – | – |
| Method | F1 | Judge | Extra calls |
|---|---|---|---|
| Scope-checked answer | 64.13 | 85.56 | 0 |
| Premise Guard | 63.70 | 84.89 | 0 |
| Coverage planning | 60.46 | 84.89 | 305 |
| w/o plan | 62.64 | 86.67 | 155 |
| w/o editor source text | 63.11 | 88.00 | 305 |
| w/o restoration | 61.09 | 86.44 | 300 |
| Contrast | PropMem | Mem0 | SimpleMem |
|---|---|---|---|
| SP Native | |||
| UP Native | |||
| SR Native | |||
| SP UP | |||
| SP SR |
| Memory system | Answer path | F1 | Judge |
|---|---|---|---|
| PropMem | Native | 62.13 | 80.72 |
| Source Reader | 51.06 | 84.09 | |
| Source + MemMend | 51.64 | 85.16 | |
| Source + ungated | 51.51 | 85.68 | |
| Mem0 | Native | 48.14 | 74.77 |
| Source Reader | 51.95 | 84.94 |
| Routing policy | Length threshold | Selected | F1 | Judge | Extra calls |
| MemMend routing (primary) | 20 | 877 | 67.17 | 88.35 | 938 |
| Without abstention branch | 20 | 368 | 64.65 | 84.56 | 429 |
| Without list branch | 20 | 681 | 66.88 | 87.78 | 741 |
| Without length branch | – | 795 | 67.14 | 88.17 | 856 |
| MemMend routing | 10 | 1,103 | 67.26 | 88.50 | 1,164 |
| MemMend routing | 30 | 813 | 67.15 | 88.25 | 874 |
| Policy | Threshold | F1 [95% CI] | Judge [95% CI] |
|---|---|---|---|
| A. Policy minus primary MemMend routing | |||
| MemMend routing (primary) | 20 | ||
| Without abstention branch | 20 | ||
| Without list branch | 20 | ||
| Without length branch | – | ||
| MemMend routing | 10 | ||
Long-Term Conversational Memory
Long-term conversational memory helps AI assistants retain, update, and use information from past conversations across sessions.
See all 292 papers