SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2606.07402 · 2026

M 3 Exam: Benchmarking Multimodal Memory for Realistic User-Agent Interactions

Wei Chen, Xiaofang Zhou, Zhoujin Tian, Fangyuan Zhang, Wenxuan Liu, Zhengjun Huang, Junle Chen, Yuqian Wu, Qintian Guo

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 19 functions out of this paper's own repositories and ran 13 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
WujiangXu/A-mem canonical 7 of 8
Mirix-AI/MIRIX canonical 3 of 7
wgcyeo/UniversalRAG canonical 3 of 3
BAI-LAB/MemoryOS — 0 of 1
FunctionStatusWhere it lives
async_cache_response Ran wgcyeo/UniversalRAG/src/universalrag/llm/openai_gpt.py
code served (permissive licence) · get_code("d00332f18998e00c")
cache_response Ran wgcyeo/UniversalRAG/src/universalrag/llm/openai_gpt.py
code served (permissive licence) · get_code("7b1455027088aa7b")
class_str Ran wgcyeo/UniversalRAG/src/universalrag/llm/openai_gpt.py
code served (permissive licence) · get_code("85d83304b08afc08")
get_field Ran Mirix-AI/MIRIX/mirix/config.py
code served (permissive licence) · get_code("b9d23ef56c0eb877")
load_config Ran Mirix-AI/MIRIX/mirix/sdk.py
code served (permissive licence) · get_code("54b0fc52f020998a")
load_locomo_dataset Ran WujiangXu/A-mem/load_dataset.py
code served (permissive licence) · get_code("5c12c257ad6b32ca")
parse_conversation Ran WujiangXu/A-mem/load_dataset.py
code served (permissive licence) · get_code("4a1510da2b417c43")
parse_session Ran WujiangXu/A-mem/load_dataset.py
code served (permissive licence) · get_code("3701fefb10719e87")
parse_with_json_fallback Ran WujiangXu/A-mem/llm_text_parsers.py
code served (permissive licence) · get_code("24e35e9dadb2fe27")
retry_llm_call Ran WujiangXu/A-mem/memory_layer_robust.py
code served (permissive licence) · get_code("ddc19d511f03d63d")
simple_tokenize Ran WujiangXu/A-mem/utils.py
code served (permissive licence) · get_code("48cea41b09d89a75")
strip_markdown_fences Ran WujiangXu/A-mem/llm_text_parsers.py
code served (permissive licence) · get_code("21854723dac05890")
validate_log_file_path Ran Mirix-AI/MIRIX/mirix/log.py
code served (permissive licence) · get_code("39746e0471f64865")
Memoryos Not yet run BAI-LAB/MemoryOS/memoryos-mcp/memoryos/memoryos.py
code served (permissive licence) · get_code("1be0d2dcd46418e9")
get_contine_chaining Not yet run Mirix-AI/MIRIX/mirix/system.py
code served (permissive licence) · get_code("67ac46c06642465e")
get_initial_boot_messages Not yet run Mirix-AI/MIRIX/mirix/system.py
code served (permissive licence) · get_code("778ddf10eda08321")
get_login_event Not yet run Mirix-AI/MIRIX/mirix/system.py
code served (permissive licence) · get_code("dec2ac6f8448f277")
parse_analyze_content Not yet run WujiangXu/A-mem/llm_text_parsers.py
code served (permissive licence) · get_code("901a2bbc34750c4e")
parse_and_chunk_text Not yet run Mirix-AI/MIRIX/mirix/embeddings.py
code served (permissive licence) · get_code("42c780d962df1807")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Language agents are increasingly deployed over accumulating multimodal information, yet existing benchmarks assume a human-human form with sparse visuals and straightforward content, evaluating neither reasoning over authentic multimodal file interaction nor the interpretation of concealed user information. We therefore introduce M 3 Exam, a query-centric multimodal conversational memory benchmark built on realistic user-agent interaction, with multi-dimensional evaluation spanning cross-modal grounding and implicit information inference. Benchmarking MLLMs and memory systems reveals persistent gaps in cross-modal grounding, cross-session reasoning, and the efficiency cost of accumulating multimodal context. We further propose M 3 Proctor, a multimodal memory method that detects query modality bias and consumes raw visual sources only on demand, improving accuracy by 13% while cutting index-construction time and retrieved tokens by over 70% 1 . * Equal contribution. 1 Code and data are released at https://anonymous. 4open.science/r/M-3-Exam-128D.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2606.07402")
get_code_for_paper("2606.07402")
have("2606.07402")

Connect an agent — have() is free.