{"id":33759,"date":"2026-06-08T08:43:43","date_gmt":"2026-06-08T15:43:43","guid":{"rendered":"https:\/\/www.pingcap.com\/?p=33759"},"modified":"2026-06-08T09:15:30","modified_gmt":"2026-06-08T16:15:30","slug":"agent-state-layer","status":"publish","type":"post","link":"https:\/\/www.pingcap.com\/ko\/blog\/agent-state-layer\/","title":{"rendered":"The Model Resets. The State Remains."},"content":{"rendered":"<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Key_Takeaways\"><\/span><strong>Key Takeaways<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Memory is what users feel. State is what engineers must get right. Most agent failures are state failures, not model failures.<\/li>\n\n\n\n<li>Memory depends on retrieval precision, not just persistence.<\/li>\n\n\n\n<li>As agents manage their own state, isolation, hybrid retrieval, and governance become product primitives.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">There is a moment using an AI agent that ruins you a little, because once you have felt it, you cannot go back. The agent remembers something from three weeks ago that you had half-forgotten you mentioned. Not in a strange way, but in a useful way. It picks up where you actually left off, not where the transcript happened to end.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Once you have had that experience a few times, going back to an agent that forgets feels like a downgrade in intelligence, even when the underlying model is the same. That is not a model capability or a UI choice. It is what happens when the system underneath is extracting, ranking, retrieving, and reconciling, deciding what to surface and what to leave quiet, in a way the user feels as continuity. What users notice first, memory, is the visible surface of something larger.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Memory_is_the_Surface_State_is_the_System\"><\/span><strong>Memory is the Surface. State is the System.<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The system underneath an AI product is not just memory. It is files, tool history, transcripts, permissions, workflow checkpoints, and facts that have to change when the world changes. That larger system is state.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">I have been thinking about this for longer than I had a product to point at. Over the last few months it stopped being a thesis and became something I could watch in production. We shipped a memory layer, and it is already in users\u2019 hands. I have watched agents create their own databases autonomously, hardware teams consolidate fragmented storage into a single source of truth, and companion-device companies treat memory as part of the product\u2019s identity. None of these teams coordinated. They arrived at the same conclusion from different starting points.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The application architectures most of us grew up with treated the database as a durable ledger behind a mostly stateless service. The application acted; the database wrote things down. Application is a verb, database is a noun. AI products invert that. The thing that acts is no longer just the application. It is a model running over state the system has selected, compressed, retrieved, and sometimes rewritten. The agent\u2019s apparent intelligence is partly what the database lets it remember and reconcile.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Memory is the layer users feel. State is the layer engineers have to get right underneath.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"A_State_Failure_Looks_Like_a_Dumb_Agent\"><\/span><strong>A State Failure Looks Like a Dumb Agent<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">When the state layer is bad, the failure rarely looks like a database problem. It looks like the agent being dumb. Here is a concrete version. An agent\u2019s memory of you, your preferences, your project context, is perfect. But the lifecycle record of what it tried and abandoned in a previous run is gone, so it starts over and repeats the same failed approach. The user experiences this as stupidity, but the agent remembered who you are. It forgot what it already tried. That is a state failure, not a memory failure, and fixing it requires a different layer of the system.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">There is a subtlety that matters even more. Memory is not about recall quantity: too little and the agent forgets the one detail that matters; too much and noise pollutes the context. What we learned from shipping a memory layer is that noise kills trust faster than forgetting does.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>The hard part of memory is not persistence. It is precision: surfacing the right information, in the right amount, at the right moment. <\/strong>Google and Anthropic reached the same conclusion from their own engineering this spring. The fix is not a bigger context window but a different architecture, one where state is explicit, durable, and decoupled from raw chat history. Latency in a state layer is no longer an SRE problem. It shows up as the agent feeling slow, stale, or like it does not quite know you.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"When_the_Databases_User_is_an_Agent\"><\/span><strong>When the Database\u2019s User is an Agent<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">I keep coming back to a number Manus told us early on. They had more than 1.2 million database clusters, and 99% had been created not by their engineers but by agents, autonomously, during runs. It is not really a scale story. It is a signal about who the database\u2019s user is becoming. If 99% of provisioning decisions are made by agents, the database is no longer managed by a DBA who plans capacity ahead of time. It is managed by a swarm of agents that do not ask. Provisioning, isolation, schema change, and TTL stop being operations tasks and become product primitives.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Dify saw a different shape of the same thing: an AI platform running nearly 500,000 database containers, one per workspace. Consolidating to a single multi-tenant system cut costs by 80% and operational overhead by 90%. The deeper lesson was not cost. It was that isolation has to become a database-native capability, not something layered on through containers.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">And agents do not only read and remember. They write files, modify configurations, and execute commands, and every operation mutates state. When something goes wrong, and in production something always does, you need to answer which agent, which run, what changed, and what the state was before. Traceable, reversible state is what makes a workspace safe for autonomous action. Without it, agents in production are a liability story, not a productivity story.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"When_the_Database_is_the_Product\"><\/span><strong>When the Database is the Product<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Plaud\u2019s version is the one I find most revealing. They make the world\u2019s leading AI voice recorder, more than 1.5 million users across 170 countries. The architecture was MySQL for metadata, S3 for transcripts. Reasonable at launch. It stops being reasonable when you have to change a schema on 300-million-row tables in a 24\/7 service, when S3 retrieval latency makes the product feel slow in the exact moment recall matters, and when partial-failure recovery becomes application logic because two systems do not share a transaction.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>For a product whose entire value proposition is remembering what was said, the database is not supporting the product. It is the product\u2019s memory. There is no clean separation.<\/strong><\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"The_Hardest_Part_is_Retrieval_Quality\"><\/span><strong>The Hardest Part is Retrieval Quality<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The hardest part, and the part that does not get talked about enough, is retrieval quality. Most teams start with vector search because semantic similarity is the natural first tool. The problem is that the scenarios where retrieval matters most are precisely where vector search alone fails. When the user mentions something specific, \u201clast Tuesday\u2019s meeting,\u201d semantic neighbors are not enough. When the answer depends on temporal structure, similarity is the wrong relation. When meaning and exact terms need to combine, you need both at once, against the same data, with the database doing the assembly.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The architectural question is whether retrieval is one query against one system or three separate round trips assembled in the application. When it is one query, with keyword match, full-text search, vector search, and SQL filtering composed into a single path, the database does the assembly. When it is three round trips, the application does, and as state grows the failure modes multiply: facts conflict, context gets noisy, and the agent starts to feel random. That randomness is fatal to trust.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">When Kimi integrated a memory layer for their agent product, we evaluated this against LoCoMo, a multi-session conversation benchmark from Snap Research. Hybrid retrieval in one query path moved recall meaningfully ahead of alternative memory systems on the same benchmark. The important part was not the number. It was that the architectural choice, one round trip rather than three, was what made the difference visible.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Bigger_Context_Windows_Wont_Fix_This\"><\/span><strong>Bigger Context Windows Won\u2019t Fix This<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The obvious counterargument is that context windows are expanding: a million tokens, two million, eventually more. If the model can see everything, why does state matter? Larger windows make the state layer more important, not less.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>A context window solves the input problem: <\/strong>What the model can see right now, in this turn.<\/li>\n\n\n\n<li><strong>State solves the management problem: <\/strong>What to keep across sessions, what to supersede, what to forget, what to govern, and what to retrieve selectively rather than dump in whole.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">The bigger the window, the more curation decisions the system has to make. A million-token window without a state layer is a million tokens of uncurated noise. The write and manage phases are where the real work happens.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"The_Fragmented_Stack_That_Breaks_at_Scale\"><\/span><strong>The Fragmented Stack That Breaks at Scale<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The teams I talk to keep hitting the same problem once their product passes prototype scale. They start with a reasonable stack: a SQL database for structured data, object storage for durability, a vector store for semantic retrieval. Each piece makes sense alone. The trouble starts when the product needs to query across all three at once, \u201cwhat did this agent remember, filtered by session, ranked by relevance, in the last seven days,\u201d and the answer has to come back in one round trip because the user is waiting.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Object storage does not coordinate transactions or isolate tenants. A vector store cannot answer which facts were rewritten and by whom. The fragmented stack works in a demo and stops working at the scale where the product\u2019s intelligence depends on querying across the fragments in real time. The team ends up rebuilding a distributed database by hand, one workaround at a time. AI workloads make this worse: they are spiky, multi-tenant, schema-fluid, and often idle until suddenly they are not.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The most interesting direction I see is agents learning to manage their own state, treating memory operations as actions they optimize rather than storage calls a developer hard-codes. The agent learns the policy; the database is what it operates on. If the substrate cannot handle the write, rewrite, branch, merge, and forget patterns at the speed that policy needs, the ceiling is infrastructure, not intelligence.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"What_AI_Products_Will_Compete_On\"><\/span><strong>What AI Products Will Compete On<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">I get asked what AI products will compete on three years from now, when everyone has roughly the same models. I do not know exactly, but I have a strong guess about what will not be enough: the model alone, and probably the UI alone. The products that pull ahead will be the ones that built a state layer that learns alongside the product, that retrieves more precisely, evolves facts more faithfully, and feels to the user like an intelligence that knows them rather than a chat that resets.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The teams making that bet early are not necessarily the most experienced. They are the ones asking a different question. Not \u201cwhat database is fast enough for our current load,\u201d but \u201cwhat does this product need to remember, govern, and evolve in two years?\u201d It feels premature to ask in the early weeks of a product. It usually is not. The model endpoint may change many times. The state your product accumulates is much harder to replace.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><em><a href=\"https:\/\/www.pingcap.com\/ko\/ai\/agentic-ai\/\">Build your AI product on a state layer designed for agents.<\/a> TiDB brings together agent-native isolation, hybrid retrieval in a single query, and lifecycle governance for production-grade agent workloads.<\/em><\/p>","protected":false},"excerpt":{"rendered":"<p>Key Takeaways There is a moment using an AI agent that ruins you a little, because once you have felt it, you cannot go back. The agent remembers something from three weeks ago that you had half-forgotten you mentioned. Not in a strange way, but in a useful way. It picks up where you actually [&hellip;]<\/p>\n","protected":false},"author":5,"featured_media":33762,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[145],"tags":[490,138,493,11,111],"class_list":["post-33759","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-thought-leadership","tag-agent-memory","tag-ai","tag-kimi","tag-real-time-analytics","tag-tidb"],"acf":[],"featured_image_src":"https:\/\/static.pingcap.com\/files\/2026\/06\/08084022\/Blog-Feature-2.png","author_info":{"display_name":"Ed Huang","author_link":"https:\/\/www.pingcap.com\/ko\/blog\/author\/ed-huang\/"},"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.6 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>The State Layer Behind Every Good AI Agent<\/title>\n<meta name=\"description\" content=\"Everyone is optimizing the model. The real differentiator is the state layer. Agent memory and consistency will decide which AI products win.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.pingcap.com\/ko\/blog\/agent-state-layer\/\" \/>\n<meta property=\"og:locale\" content=\"ko_KR\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"The State Layer Behind Every Good AI Agent\" \/>\n<meta property=\"og:description\" content=\"Everyone is optimizing the model. The real differentiator is the state layer. Agent memory and consistency will decide which AI products win.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.pingcap.com\/ko\/blog\/agent-state-layer\/\" \/>\n<meta property=\"og:site_name\" content=\"TiDB\" \/>\n<meta property=\"article:publisher\" content=\"https:\/\/facebook.com\/pingcap2015\" \/>\n<meta property=\"article:published_time\" content=\"2026-06-08T15:43:43+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2026-06-08T16:15:30+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/static.pingcap.com\/files\/2026\/06\/08084002\/Blog-Feature-Banner-3.png\" \/>\n\t<meta property=\"og:image:width\" content=\"1600\" \/>\n\t<meta property=\"og:image:height\" content=\"900\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"author\" content=\"Ed Huang\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@PingCAP\" \/>\n<meta name=\"twitter:site\" content=\"@PingCAP\" \/>\n<meta name=\"twitter:label1\" content=\"\uae00\uc4f4\uc774\" \/>\n\t<meta name=\"twitter:data1\" content=\"Ed Huang\" \/>\n\t<meta name=\"twitter:label2\" content=\"\uc608\uc0c1 \ub418\ub294 \ud310\ub3c5 \uc2dc\uac04\" \/>\n\t<meta name=\"twitter:data2\" content=\"9\ubd84\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/agent-state-layer\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/agent-state-layer\\\/\"},\"author\":{\"name\":\"Ed Huang\",\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/#\\\/schema\\\/person\\\/e2787ee59fb58fadf52912ddc6e0c7ef\"},\"headline\":\"The Model Resets. The State Remains.\",\"datePublished\":\"2026-06-08T15:43:43+00:00\",\"dateModified\":\"2026-06-08T16:15:30+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/agent-state-layer\\\/\"},\"wordCount\":1851,\"publisher\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/#organization\"},\"image\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/agent-state-layer\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/static.pingcap.com\\\/files\\\/2026\\\/06\\\/08084022\\\/Blog-Feature-2.png\",\"keywords\":[\"Agent Memory\",\"AI\",\"Kimi\",\"Real-time analytics\",\"TiDB\"],\"articleSection\":[\"Thought Leadership\"],\"inLanguage\":\"ko-KR\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/agent-state-layer\\\/\",\"url\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/agent-state-layer\\\/\",\"name\":\"The State Layer Behind Every Good AI Agent\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/agent-state-layer\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/agent-state-layer\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/static.pingcap.com\\\/files\\\/2026\\\/06\\\/08084022\\\/Blog-Feature-2.png\",\"datePublished\":\"2026-06-08T15:43:43+00:00\",\"dateModified\":\"2026-06-08T16:15:30+00:00\",\"description\":\"Everyone is optimizing the model. The real differentiator is the state layer. Agent memory and consistency will decide which AI products win.\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/agent-state-layer\\\/#breadcrumb\"},\"inLanguage\":\"ko-KR\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/agent-state-layer\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"ko-KR\",\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/agent-state-layer\\\/#primaryimage\",\"url\":\"https:\\\/\\\/static.pingcap.com\\\/files\\\/2026\\\/06\\\/08084022\\\/Blog-Feature-2.png\",\"contentUrl\":\"https:\\\/\\\/static.pingcap.com\\\/files\\\/2026\\\/06\\\/08084022\\\/Blog-Feature-2.png\",\"width\":1800,\"height\":600},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/agent-state-layer\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/www.pingcap.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"The Model Resets. The State Remains.\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/#website\",\"url\":\"https:\\\/\\\/www.pingcap.com\\\/\",\"name\":\"TiDB\",\"description\":\"TiDB | SQL at Scale\",\"publisher\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/www.pingcap.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"ko-KR\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/#organization\",\"name\":\"PingCAP\",\"url\":\"https:\\\/\\\/www.pingcap.com\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"ko-KR\",\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/static.pingcap.com\\\/files\\\/2021\\\/11\\\/pingcap-logo.png\",\"contentUrl\":\"https:\\\/\\\/static.pingcap.com\\\/files\\\/2021\\\/11\\\/pingcap-logo.png\",\"width\":811,\"height\":232,\"caption\":\"PingCAP\"},\"image\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/#\\\/schema\\\/logo\\\/image\\\/\"},\"sameAs\":[\"https:\\\/\\\/facebook.com\\\/pingcap2015\",\"https:\\\/\\\/x.com\\\/PingCAP\",\"https:\\\/\\\/linkedin.com\\\/company\\\/pingcap\",\"https:\\\/\\\/youtube.com\\\/channel\\\/UCuq4puT32DzHKT5rU1IZpIA\"]},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/#\\\/schema\\\/person\\\/e2787ee59fb58fadf52912ddc6e0c7ef\",\"name\":\"Ed Huang\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"ko-KR\",\"@id\":\"https:\\\/\\\/static.pingcap.com\\\/files\\\/2022\\\/10\\\/17234942\\\/avatar.jpg\",\"url\":\"https:\\\/\\\/static.pingcap.com\\\/files\\\/2022\\\/10\\\/17234942\\\/avatar.jpg\",\"contentUrl\":\"https:\\\/\\\/static.pingcap.com\\\/files\\\/2022\\\/10\\\/17234942\\\/avatar.jpg\",\"caption\":\"Ed Huang\"},\"description\":\"Co-Founder and CTO\",\"sameAs\":[\"https:\\\/\\\/www.linkedin.com\\\/in\\\/eddxhuang\\\/\"],\"url\":\"https:\\\/\\\/www.pingcap.com\\\/ko\\\/blog\\\/author\\\/ed-huang\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"The State Layer Behind Every Good AI Agent","description":"Everyone is optimizing the model. The real differentiator is the state layer. Agent memory and consistency will decide which AI products win.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.pingcap.com\/ko\/blog\/agent-state-layer\/","og_locale":"ko_KR","og_type":"article","og_title":"The State Layer Behind Every Good AI Agent","og_description":"Everyone is optimizing the model. The real differentiator is the state layer. Agent memory and consistency will decide which AI products win.","og_url":"https:\/\/www.pingcap.com\/ko\/blog\/agent-state-layer\/","og_site_name":"TiDB","article_publisher":"https:\/\/facebook.com\/pingcap2015","article_published_time":"2026-06-08T15:43:43+00:00","article_modified_time":"2026-06-08T16:15:30+00:00","og_image":[{"width":1600,"height":900,"url":"https:\/\/static.pingcap.com\/files\/2026\/06\/08084002\/Blog-Feature-Banner-3.png","type":"image\/png"}],"author":"Ed Huang","twitter_card":"summary_large_image","twitter_creator":"@PingCAP","twitter_site":"@PingCAP","twitter_misc":{"\uae00\uc4f4\uc774":"Ed Huang","\uc608\uc0c1 \ub418\ub294 \ud310\ub3c5 \uc2dc\uac04":"9\ubd84"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.pingcap.com\/blog\/agent-state-layer\/#article","isPartOf":{"@id":"https:\/\/www.pingcap.com\/blog\/agent-state-layer\/"},"author":{"name":"Ed Huang","@id":"https:\/\/www.pingcap.com\/#\/schema\/person\/e2787ee59fb58fadf52912ddc6e0c7ef"},"headline":"The Model Resets. The State Remains.","datePublished":"2026-06-08T15:43:43+00:00","dateModified":"2026-06-08T16:15:30+00:00","mainEntityOfPage":{"@id":"https:\/\/www.pingcap.com\/blog\/agent-state-layer\/"},"wordCount":1851,"publisher":{"@id":"https:\/\/www.pingcap.com\/#organization"},"image":{"@id":"https:\/\/www.pingcap.com\/blog\/agent-state-layer\/#primaryimage"},"thumbnailUrl":"https:\/\/static.pingcap.com\/files\/2026\/06\/08084022\/Blog-Feature-2.png","keywords":["Agent Memory","AI","Kimi","Real-time analytics","TiDB"],"articleSection":["Thought Leadership"],"inLanguage":"ko-KR"},{"@type":"WebPage","@id":"https:\/\/www.pingcap.com\/blog\/agent-state-layer\/","url":"https:\/\/www.pingcap.com\/blog\/agent-state-layer\/","name":"The State Layer Behind Every Good AI Agent","isPartOf":{"@id":"https:\/\/www.pingcap.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.pingcap.com\/blog\/agent-state-layer\/#primaryimage"},"image":{"@id":"https:\/\/www.pingcap.com\/blog\/agent-state-layer\/#primaryimage"},"thumbnailUrl":"https:\/\/static.pingcap.com\/files\/2026\/06\/08084022\/Blog-Feature-2.png","datePublished":"2026-06-08T15:43:43+00:00","dateModified":"2026-06-08T16:15:30+00:00","description":"Everyone is optimizing the model. The real differentiator is the state layer. Agent memory and consistency will decide which AI products win.","breadcrumb":{"@id":"https:\/\/www.pingcap.com\/blog\/agent-state-layer\/#breadcrumb"},"inLanguage":"ko-KR","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.pingcap.com\/blog\/agent-state-layer\/"]}]},{"@type":"ImageObject","inLanguage":"ko-KR","@id":"https:\/\/www.pingcap.com\/blog\/agent-state-layer\/#primaryimage","url":"https:\/\/static.pingcap.com\/files\/2026\/06\/08084022\/Blog-Feature-2.png","contentUrl":"https:\/\/static.pingcap.com\/files\/2026\/06\/08084022\/Blog-Feature-2.png","width":1800,"height":600},{"@type":"BreadcrumbList","@id":"https:\/\/www.pingcap.com\/blog\/agent-state-layer\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.pingcap.com\/"},{"@type":"ListItem","position":2,"name":"The Model Resets. The State Remains."}]},{"@type":"WebSite","@id":"https:\/\/www.pingcap.com\/#website","url":"https:\/\/www.pingcap.com\/","name":"\ud2f0DB","description":"TiDB | SQL at Scale","publisher":{"@id":"https:\/\/www.pingcap.com\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.pingcap.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"ko-KR"},{"@type":"Organization","@id":"https:\/\/www.pingcap.com\/#organization","name":"PingCAP","url":"https:\/\/www.pingcap.com\/","logo":{"@type":"ImageObject","inLanguage":"ko-KR","@id":"https:\/\/www.pingcap.com\/#\/schema\/logo\/image\/","url":"https:\/\/static.pingcap.com\/files\/2021\/11\/pingcap-logo.png","contentUrl":"https:\/\/static.pingcap.com\/files\/2021\/11\/pingcap-logo.png","width":811,"height":232,"caption":"PingCAP"},"image":{"@id":"https:\/\/www.pingcap.com\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/facebook.com\/pingcap2015","https:\/\/x.com\/PingCAP","https:\/\/linkedin.com\/company\/pingcap","https:\/\/youtube.com\/channel\/UCuq4puT32DzHKT5rU1IZpIA"]},{"@type":"Person","@id":"https:\/\/www.pingcap.com\/#\/schema\/person\/e2787ee59fb58fadf52912ddc6e0c7ef","name":"Ed Huang","image":{"@type":"ImageObject","inLanguage":"ko-KR","@id":"https:\/\/static.pingcap.com\/files\/2022\/10\/17234942\/avatar.jpg","url":"https:\/\/static.pingcap.com\/files\/2022\/10\/17234942\/avatar.jpg","contentUrl":"https:\/\/static.pingcap.com\/files\/2022\/10\/17234942\/avatar.jpg","caption":"Ed Huang"},"description":"Co-Founder and CTO","sameAs":["https:\/\/www.linkedin.com\/in\/eddxhuang\/"],"url":"https:\/\/www.pingcap.com\/ko\/blog\/author\/ed-huang\/"}]}},"grav_blocks":false,"card_markup":"<a class=\"card-resource bg-white\" href=\"https:\/\/www.pingcap.com\/ko\/blog\/agent-state-layer\/\"><div class=\"card-resource__image-container\"><img class=\"card-resource__image\" alt=\"Blog - Feature-2\" src=\"https:\/\/static.pingcap.com\/files\/2026\/06\/08084022\/Blog-Feature-2.png\" loading=\"lazy\" width=1800 height=600 \/><\/div><div class=\"card-resource__content-container\"><div class=\"card-resource__content-head\"><div class=\"card-resource__category\">Thought Leadership<\/div><\/div><h5 class=\"card-resource__title\">The Model Resets. The State Remains.<\/h5><\/div><\/a>","_links":{"self":[{"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/posts\/33759","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/users\/5"}],"replies":[{"embeddable":true,"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/comments?post=33759"}],"version-history":[{"count":4,"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/posts\/33759\/revisions"}],"predecessor-version":[{"id":33766,"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/posts\/33759\/revisions\/33766"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/media\/33762"}],"wp:attachment":[{"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/media?parent=33759"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/categories?post=33759"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/tags?post=33759"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}