[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85731-en":3,"doc-seo-85731-105":29,"detail-sidebar-cat-0-en-105":90},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},85731,4398048950312,"Violet","https://ap-avatar.wpscdn.com/avatar/400002538284de19e3c?_k=1778320343897328908",8,"Research & Report","Memory-Conditioned Tool Calling for Camera-First Visual Agents","Personal memory influences what an image-focused agent should look up next, yet camera-first interaction restricts users to sending only images, forcing the agent to decide tool calls autonomously. This work studies how personal visual memory improves agent-side tool choice and tool arguments, producing more user-aligned multi-tool lookups. A three-layer memory (profile, short-term focus, observations) conditions an LLM tool-calling loop per turn, with conflict-aware write-back for later captures. On 800 images with synthetic memory ablations, removing the full memory block lowers tool-query relevance and end-to-end utility, isolating memory conditioning under fixed synthetic blocks.","arXiv :2607 .09822v1 [ cs .CV] 10 Jul 2026  \nMemory-Conditioned Tool Calling for Camera-First Visual Agents  \nXiaofan Wu∗1, Xi Zeng 1 , Miaoxia Chen 1 , Peishan Chen 1 , Shuyan Li 1 , Jiyun Yao 1 ,  \nHanyong Zhong 1 , Jiahao Zhu 1  \n1 Chance AI  \nJuly 2026  \nAbstract  \nRecognition tells an agent what is in an image; personal memory affects what is worth looking up next. In a camera-first setting the user can send only an image, so the agent must form the lookups. We study whether personal visual memory improves agent-side tool choice and tool arguments, and thereby more user-aligned multi-tool lookups. The design uses a three-layer personal visual memory (profile, short-term focus, observations) that is loaded on each turn to condition an LLM tool-calling loop under camera-first intake, and includes conflict-aware write-back intended to refresh the user model for later captures. On 800 images paired with synthetic memory blocks constructed for controlled ablation, removing the full three-layer memory block reduces tool-query relevance by 0.47 points absolute (4.21 → 3.74 on a 5-point scale; 11.2% relative) and end-to-end utility by 0.082 absolute (0.842 → 0.760; 9.7% relative) . These results measure memory conditioning of tool policy under image-only intake with fixed synthetic blocks, not multi-session write-back from live user histories.  \n1 Introduction  \nLooking at the same garment, a fashion designer and a software engineer tend to look up different things: knowledge level, taste, past comparisons, and the next decision all shape what is worth retrieving. Visual systems are often judged by recognition—labels, matches,“what is it?”—but recognition alone does not decide which follow-up lookups to make.  \nThis paper measures one link in that process for camera-first agents, where the user can send only an image and the agent must form tool calls itself:  \npersonal memory → better agent tool calls → more user-aligned multi-tool lookups.  \nA tool call is which tool to invoke and with what arguments. Better means more relevant, more specific, and better matched to the user. With memory, the agent issues more focused tool calls; focused tool calls support more useful lookups; the design also includes write-back so later images can load an updated memory block (not measured as multi-session compounding here) .  \nMany multimodal agents [2, 3 , 4] search without a persistent user model. When there is no typed query, tool calls stay generic even if recognition is correct. Personal memory is one way to condition those tool calls on the user.  \nWe study memory-conditioned tool calling for camera-first agents: personal memory is recalled on each turn and conditions tool choice and arguments; write-back is part of the system design for later captures. The design implements the link as follows:  \n∗ Corresponding author: [xiaofan@chance.vision](xiaofan@chance.vision).  \n• Memory substrate. Three layers—profile, short-term focus, and observations—encode who the user is and what they are currently exploring. Full memory means all three layers are present.  \n• Inner loop (this turn) . Memory conditions tool choice and arguments; results may refine the next tool call within the same turn.  \n• Outer write-back (design; next capture) . Background updates are intended to refresh observations and, over time, profile and short-term focus for later images; this paper does not measure multi-session compounding from write-back.  \nContributions. (i) We treat memory-conditioned tool policy as the interface where personal visual memory meets camera-first multi-tool lookup, complementary to dialogue-memory work on multi-session factual recall [13, 1] . (ii) We describe a three-layer memory design (profile, short-term focus, observations) with conflict-aware observation write-back and a multi-tool loop suitable for image-only intake; write-back is a design component, not an evaluated outcome here.  \n(iii) Under controlled synthetic memory blocks (not liv","cbCaic60I6UCB5J4","https://ap.wps.com/l/cbCaic60I6UCB5J4","pdf",457047,1,14,"English","en",105,"# Abstract\n# Introduction\n# Related Work\n## Multimodal Search Agents\n## Retrieval-Augmented Vision-Language Models\n## Tool-Using Language Agents\n## Memory and Personalization","[{\"question\":\"What problem does the paper target in camera-first visual agents?\",\"answer\":\"In camera-first settings, users provide only an image, so the agent must autonomously form tool calls and determine what to retrieve next. The paper targets how personal visual memory can make those tool calls more user-aligned.\"},{\"question\":\"How is personal visual memory modeled and used during tool calling?\",\"answer\":\"The system uses a three-layer memory—profile, short-term focus, and observations—loaded at each turn. This memory conditions both tool choice and tool arguments inside the LLM tool-calling loop.\"},{\"question\":\"What experimental evidence shows the effect of memory conditioning?\",\"answer\":\"On 800 images with controlled synthetic memory blocks, ablating the full three-layer memory block reduces tool-query relevance (4.21 to 3.74 on a 5-point scale) and decreases end-to-end utility (0.842 to 0.760). The study isolates memory conditioning while keeping visual context, tools, and model fixed.\"}]",1784205871,35,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":85,"head_meta":87,"extra_data":89,"updated_unix":27},"memory-conditioned-tool-calling-for-camera-first-visual-agents","",{"@graph":35,"@context":84},[36,53,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/memory-conditioned-tool-calling-for-camera-first-visual-agents/85731/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":61,"encodingFormat":60,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":4},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"What problem does the paper target in camera-first visual agents?","Question",{"text":74,"@type":75},"In camera-first settings, users provide only an image, so the agent must autonomously form tool calls and determine what to retrieve next. The paper targets how personal visual memory can make those tool calls more user-aligned.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How is personal visual memory modeled and used during tool calling?",{"text":79,"@type":75},"The system uses a three-layer memory—profile, short-term focus, and observations—loaded at each turn. This memory conditions both tool choice and tool arguments inside the LLM tool-calling loop.",{"name":81,"@type":72,"acceptedAnswer":82},"What experimental evidence shows the effect of memory conditioning?",{"text":83,"@type":75},"On 800 images with controlled synthetic memory blocks, ablating the full three-layer memory block reduces tool-query relevance (4.21 to 3.74 on a 5-point scale) and decreases end-to-end utility (0.842 to 0.760). The study isolates memory conditioning while keeping visual context, tools, and model fixed.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,119,122,127,130,134],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":105,"slug":137},19,"General","general"]