MemDoctor: Stage-wise Error Attribution and Repair for MLLM Agent Memory
Organizations: Shanghai Jiao Tong University · Xi’an Jiaotong University · Xidian University
Abstract
Multimodal large language models (MLLMs) now assist users across sessions of images and text, carried by an external memory of what was said and shown. When such an assistant answers wrongly, the failure usually began earlier, in how a record was written, updated or retrieved, and the answer alone does not say where. We formalize multimodal memory as four stages, Write, Update, Retrieve and Answer, and define nine error types across them. A stage-resolved analysis on Mem-Gallery traces most failures to a source stage, and most of those lie before the answer. Injecting known errors against fact-preserving control edits, we then follow each tainted record downstream. Early errors survive to the answer as apparent evidence, and hurt most when no other record holds the fact. A diagnosis that trusts the quoted evidence therefore blames the stage where the error surfaces, not the one where it began. We propose MemDoctor, in which the MLLM that built the memory diagnoses its own failures. It traces each needed fact backward through the four stages, accepts a diagnosis only when the cited evidence is found in memory, and repairs that stage before answering again. On a 4B backbone, MemDoctor improves accuracy on both Mem-Gallery and MemLens, outperforming all evaluated 8B baselines, and the same loop also improves MuRAG and A-Mem.
Figures & tables
| Stage | Faulty when | Error types | Repair |
|---|---|---|---|
| Write | omission, distortion, visual loss, visual misdescription | rewrite from | |
| Update | or buries | update loss | re-merge from versions and |
| Retrieve | or derails | retrieval miss, retrieval distraction | re-query |
| Answer | misreasoning, false recall | re-answer from evidence | |
| None | — | no change |
| Method | FR | MR | TR | VR | TTL | VS | CD | KR | AR | Overall |
|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-VL-8B-Instruct Backbone | ||||||||||
| TextRAG | 70.78 | 74.27 | 66.67 | 41.95 | 75.67 | 56.54 | 66.67 | 64.20 | 94.57 | 68.44 |
| MuRAG | 71.69 | 78.16 | 69.92 | 49.43 | 85.76 | 75.82 | 62.96 | 62.96 | 94.02 | 75.16 |
| A-Mem | 68.49 | 61.17 | 69.92 | 55.17 | 64.69 | 67.97 | 32.10 | 67.90 | 96.74 | 66.80 |
| Mem0 | 70.32 | 58.74 | 33.33 | 40.80 | 67.66 | 8.50 | 33.33 | 54.32 | 97.28 | 52.07 |
| Omni-SimpleMem | 71.23 | 62.14 | 71.54 | 57.47 | 69.44 | 68.95 | 34.57 | 61.73 | 93.48 | 68.21 |
| Tainted record | |||||||
|---|---|---|---|---|---|---|---|
| Stage | Error | in memory | retrieved | quoted | Wrong answer (inj. / ctrl.) | [95% CI] | |
| Write | Omission | 130 | 65% | 85% | 62% | 28% / 3% | +25 [+19, +34] |
| Write | Distortion | 133 | 80% | 100% | 97% | 45% / 2% | +43 [+35, +53] |
| Write | Visual loss | 12 | 50% | 100% | 67% | 50% / 0% | +50 [+25, +75] |
| Write | Visual misdescription | 12 | 17% | 100% | 0% | 33% / 0% | +33 [0, +59] |
| Update | Update loss | 11 | 73% | 91% | 80% | 45% / 0% | +45 [+17, +80] |
| Method | IE | MSR | TR | KU | AR | Overall | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-VL-8B-Instruct Backbone | ||||||||||||
| TextRAG | 42.6 | 17.1 | 50.0 | 34.5 | 100.0 | 45.1 | ||||||
| MuRAG | 37.7 | 17.1 | 52.1 | 27.6 | 100.0 | 43.1 | ||||||
| A-Mem | 42.6 | 11.4 | 47.9 | 31.0 | 100.0 | 43.1 | ||||||
| Mem0 | 18.0 | 22.9 | 39.6 | 34.5 | 100.0 | 35.9 | ||||||
| Omni-SimpleMem | 37.7 | 14.3 | 43.8 | 27.6 | 100.0 | 40.5 | ||||||
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
| Error | Redundant | Wrong answer (inj. / ctrl.) | ||
|---|---|---|---|---|
| Omission | no | 49 | 61% / 2% | +59 |
| yes | 81 | 9% / 4% | +5 | |
| Distortion | no | 75 | 71% / 3% | +68 |
| yes | 58 | 12% / 2% | +10 | |
| Visual loss | no | 10 | 40% / 0% | +40 |
| yes | 2 | 100% / 0% | +100 |
| Procedure | IE (61) | KU (29) | TR (48) | MSR (35) | AR (22) | Overall (195) | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MuRAG | ||||||||||||
| Original | 39.34 | 31.03 | 52.08 | 17.14 | 100.00 | 44.10 | ||||||
| Reflexion | 50.82 | +11.48 | 37.93 | +6.90 | 52.08 | 0.00 | 17.14 | 0.00 | 100.00 | 0.00 | 48.72 | +4.62 |
| MemDoctor | 59.02 | +19.67 | 37.93 | +6.90 | 54.17 | +2.08 | 17.14 | 0.00 | 100.00 | 0.00 | 51.79 | +7.69 |
| A-Mem | ||||||||||||
| Original | 44.26 | 31.03 | 45.83 | 14.29 | 100.00 | 43.59 | ||||||
| Procedure | FR | MR | TR | VR | TTL | VS | CD | KR | AR | Overall | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MuRAG | ||||||||||||||||||||
| Original | 71.69 | 78.16 | 69.92 | 49.43 | 85.76 | 75.82 | 62.96 | 62.96 | 94.02 | 75.16 | ||||||||||
| Reflexion | 72.15 | 80.10 | 74.80 | 57.47 | 85.16 | 77.12 | 65.43 | 62.96 | 86.96 | 76.10 | ||||||||||
| MemDoctor | 73.52 | +1.83 | 78.64 | +0.49 | 73.17 | +3.25 | 51.72 | +2.30 | 85.46 | 0.30 | 78.10 | +2.29 | 66.67 | +3.70 | 66.67 | +3.70 | 94.02 | 0.00 | 76.62 | +1.46 |
| A-Mem | ||||||||||||||||||||
| Original | 69.41 | 60.68 | 68.29 | 51.72 | 63.20 | 66.01 | 32.10 | 71.60 | 95.11 | 65.75 | ||||||||||
LLM Agent Memory
Large language model agent memory lets agents retain, organize, retrieve, and update information across interactions and long-running tasks.
See all 399 papers