{"id":33693,"date":"2026-06-03T13:33:20","date_gmt":"2026-06-03T20:33:20","guid":{"rendered":"https:\/\/www.pingcap.com\/?p=33693"},"modified":"2026-06-05T17:16:28","modified_gmt":"2026-06-06T00:16:28","slug":"ai-agent-memory-runtime-production-agents","status":"publish","type":"post","link":"https:\/\/www.pingcap.com\/ko\/blog\/ai-agent-memory-runtime-production-agents\/","title":{"rendered":"Beyond Context Windows: Memory is the Runtime for Production Agents"},"content":{"rendered":"<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<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>Context windows are not memory. Production agents need durable systems for persistence, updates, expiration, and deletion.<\/li>\n\n\n\n<li>Agent memory has two sides: cognitive context and workspace evidence.<\/li>\n\n\n\n<li>Memory must be designed around type, scope, and time.<\/li>\n\n\n\n<li>Reliable memory requires recall, forgetting, consistency, observability, and policy controls.<\/li>\n\n\n\n<li>mem9, drive9, and TiDB Cloud unify agent memory, workspace state, retrieval, governance, and scale.<\/li>\n<\/ul>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">Ask a production agent a simple operational question and the cracks show immediately. What did this agent know last Tuesday? Which facts were true before a customer changed their policy? Which records must we purge everywhere because they contain sensitive data, including copies in a semantic index and a restored snapshot? Most agent stacks today cannot answer any of these questions because they still treat \u201cAI agent memory\u201d as a chat transcript embedded into a vector store and dropped back into a prompt.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">That was enough when an agent was a conversational wrapper around a model. That approach does not support agents that run across sessions, collaborate with other agents, act on live business data, and need to remember correctly without becoming impossible to govern. Once an agent crosses that threshold, memory stops behaving like a convenience feature and starts behaving like runtime infrastructure. The production memory runtime has two sides: Cognitive memory for durable agent context, and workspace memory for the files, artifacts, and evidence around real work. Our answer to both is a pair of systems, <a href=\"https:\/\/mem9.ai\/\">mem9<\/a> \uadf8\ub9ac\uace0 <a href=\"https:\/\/drive9.ai\/\">drive9<\/a>, built on one operationally sound foundation.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"AI_Agent_Memory_From_Retrieval_Add-On_to_Runtime_Layer\"><\/span>AI Agent Memory: <strong>From Retrieval Add-On to Runtime Layer<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A context window is a transport mechanism, not a memory strategy. A context window can carry information into a single model call, but it cannot decide what to persist, update, expire, or delete everywhere.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/www.linkedin.com\/blog\/engineering\/ai\/the-linkedin-generative-ai-application-tech-stack-personalization-with-cognitive-memory-agent\">LinkedIn Engineering recently published<\/a> one of the clearest public write-ups of why memory now belongs inside the application stack. Their taxonomy (conversational, episodic, semantic, procedural) is a good starting point, and the more important point is structural: Retrieval is a reasoning process, not a single search operation. Once memory spans both cognitive state and workspace evidence, recall has to reason across type, scope, and time.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The distinction between transport and runtime sounds semantic until the first operational question arrives. Debugging, evaluation, compliance, and undo all require asking what the agent knew at a specific moment, not only what it knows now. Memory that cannot answer those questions is a retrieval add-on. Memory that can is a runtime with lifecycle rules.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Type_Scope_and_Time_the_Real_Design_Unit\"><\/span><strong>Type, Scope, and Time: the Real Design Unit<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The layered memory model is a useful start, but it is not enough on its own. In production, the design unit is type by scope by time. Type tells you what kind of memory you are storing. Scope tells you who or what that memory belongs to. Time tells you whether it is current, revisable, expired, or recallable at a specific moment.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">These three axes operate at once. A user preference (cognitive type, user scope) stays authoritative until the user changes it. A project decision (cognitive type, project scope) is durable but revisable. An operational trace (workspace type, ops scope) captures point-in-time evidence that no one should silently rewrite. Forcing all of these into one undifferentiated memory bucket usually weakens one side.<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><td><strong>Type<\/strong><\/td><td><strong>User<\/strong><\/td><td><strong>Task<\/strong><\/td><td><strong>Project<\/strong><\/td><td><strong>Ops<\/strong><\/td><\/tr><tr><td>Cognitive<\/td><td>Preferences, reusable instructions<\/td><td>Intent, open questions, next steps<\/td><td>Decisions, conventions, architecture assumptions<\/td><td>Policies, operating rules<\/td><\/tr><tr><td>Workspace<\/td><td>Uploads, personal files<\/td><td>Working state, briefs, notes, handoffs, task artifacts<\/td><td>Design docs, architecture notes, roadmaps, decision logs<\/td><td>Traces, runbooks, audit evidence<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Time applies across every cell. Every memory item carries one of three states, current, durable, or point-in-time, and each state keeps the item recallable, rollback-ready, and auditable exactly as it existed at that moment.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><em>Note: Task scope is short-lived memory for one active unit of work. Project scope is longer-lived shared memory that survives across tasks. Keeping them separate prevents temporary working records from being treated as durable project knowledge and state.<\/em><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This split is exactly where mem9 and drive9 complement each other. mem9 is the cognitive side: Session traces for ongoing interaction state, extracted insights for semantic facts, and pinned memories for durable curated context. drive9 is the workspace side: Files, revisions, extracted text, descriptions, and task state. Preferences, reusable instructions, and project decisions want the reconciliation and compact memory forms mem9 provides. Logs, runbooks, design docs, traces, and code want the path semantics, revisions, and provenance drive9 provides. Together they give an agent both a memory of what it has learned and a workspace of what it can reopen, verify, and reuse.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Ingestion_Decides_What_Becomes_AI_Agent_Memory\"><\/span><strong>Ingestion Decides What Becomes AI Agent Memory<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Not every input should be stored the same way, and good ingestion is a decision, not a parsing step. The core question is whether an input should be rewritten into compact memory or preserved as inspectable evidence.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">On the cognitive side, mem9 ingestion captures raw turns inline, extracts candidate facts, and reconciles them against existing memory. Later passes summarize, revise, or promote what should remain durable. On the workspace side, drive9 ingestion commits the file, path, revision, and immediately available text inline, and later workers enrich that artifact with descriptions, media extraction, or embeddings when useful. One path rewrites information into memory. The other preserves the source artifact as evidence. The agent needs both, and it needs to know which is which.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"559\" src=\"https:\/\/static.pingcap.com\/files\/2026\/06\/04124932\/image-1-1024x559.png\" alt=\"A sample AI agent memory architecture featuring mem9 and drive9.\" class=\"wp-image-33695\" srcset=\"https:\/\/static.pingcap.com\/files\/2026\/06\/04124932\/image-1-1024x559.png 1024w, https:\/\/static.pingcap.com\/files\/2026\/06\/04124932\/image-1-300x164.png 300w, https:\/\/static.pingcap.com\/files\/2026\/06\/04124932\/image-1-768x419.png 768w, https:\/\/static.pingcap.com\/files\/2026\/06\/04124932\/image-1-1536x838.png 1536w, https:\/\/static.pingcap.com\/files\/2026\/06\/04124932\/image-1-18x10.png 18w, https:\/\/static.pingcap.com\/files\/2026\/06\/04124932\/image-1.png 2048w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"If_AI_Agent_Memory_Is_Infrastructure_It_Needs_Infrastructure_Properties\"><\/span><strong>If AI Agent Memory Is Infrastructure, It Needs Infrastructure Properties<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Once memory becomes shared, multi-session, and business-relevant, the problem stops being &#8220;how do I retrieve a relevant note?&#8221; and becomes &#8220;how do I run stateful systems safely across cognitive state and workspace state?&#8221; Four properties separate a memory runtime from a retrieval feature.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Point-in-time recall.<\/strong> The ability to ask what the agent knew at a specific moment, not only what it knows now. Essential for debugging, evaluation, compliance, and undo.<\/li>\n\n\n\n<li><strong>Selective forgetting.<\/strong> Deleting a record is not enough if the same content survives in a semantic index, a derived summary, a file revision, or a restored snapshot. Memory systems need purge semantics, meaning a delete that propagates everywhere, not just a delete button.<\/li>\n\n\n\n<li><strong>Consistency boundaries.<\/strong> When an agent changes a project state, stores a memory in mem9, and writes an artifact to drive9, those operations cannot drift apart. Durable behavior needs clear consistency boundaries and revision semantics.<\/li>\n\n\n\n<li><strong>Observability and policy.<\/strong> Lineage, auditability, retention, and scope-aware access controls. The memory layer becomes part of the control plane, not just the prompt pipeline.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Once state, files, time, and policy interact this way, the stack below the memory API stops looking like a single retrieval service and starts looking like a database-backed application runtime.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"A_Pragmatic_Stack_TiDB_Cloud_mem9_and_drive9\"><\/span><strong>A Pragmatic Stack: TiDB Cloud, mem9, and drive9<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A practical way to think about the stack is to separate cognitive memory from workspace memory, then keep both on one consolidated data substrate. That is what mem9, drive9, and <a href=\"https:\/\/www.pingcap.com\/ko\/tidb\/cloud\/?utm_source=chatgpt.com\">TiDB Cloud<\/a> are together.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>mem9 is cognitive memory.<\/strong> It is the system for compact, revisable agent context: Session traces, extracted insights, and durable pinned knowledge that can be carried across runs, shared between agents, and updated as the system learns.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>drive9 is workspace memory.<\/strong> It gives agents one path namespace over files, revisions, extracted text, descriptions, and task state, so valuable workspace changes do not disappear with an ephemeral sandbox. Every file, configuration, and artifact mutation can be versioned, audited, and restored across runtimes such as E2B, Daytona, local containers, or other agent environments. drive9 cares less about what the agent knows and more about what evidence the agent can reopen, verify, restore, and reuse.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>TiDB Cloud is the managed distributed database foundation<\/strong> underneath both. It provides the scale, consistency, retrieval, and isolation production agents need, giving mem9 and drive9 one durable backend for memory, task history, metadata, files, artifacts, and operational records. mem9 gives agents memory; drive9 gives agents persistent workspace state; TiDB Cloud keeps both reliable, searchable, and production-ready.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This pattern is broader than our own stack. Memori Labs, which builds agent-native persistent memory for LLMs and <a href=\"https:\/\/memorilabs.ai\/docs\/memori-byodb\/databases\/tidb\/\">supports TiDB as a BYODB backend<\/a>, hits the same production requirements described here: Memory scoped by user, agent, and session; stored in the customer&#8217;s own infrastructure; transformed asynchronously into structured facts, preferences, embeddings, and knowledge graph primitives; and recalled across future agent runs, including execution traces like tool calls, decisions, and failures. Memori signals the same shift: production agent memory is structured, persistent, attributable state, not vector search over chat history.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"540\" src=\"https:\/\/static.pingcap.com\/files\/2026\/06\/04124929\/image-1024x540.png\" alt=\"Memori Labs' reference architecture featuring TiDB.\" class=\"wp-image-33694\" srcset=\"https:\/\/static.pingcap.com\/files\/2026\/06\/04124929\/image-1024x540.png 1024w, https:\/\/static.pingcap.com\/files\/2026\/06\/04124929\/image-300x158.png 300w, https:\/\/static.pingcap.com\/files\/2026\/06\/04124929\/image-768x405.png 768w, https:\/\/static.pingcap.com\/files\/2026\/06\/04124929\/image-1536x809.png 1536w, https:\/\/static.pingcap.com\/files\/2026\/06\/04124929\/image-18x9.png 18w, https:\/\/static.pingcap.com\/files\/2026\/06\/04124929\/image.png 2048w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">That is the strongest argument for keeping mem9 and drive9 on TiDB and TiKV, and it is not &#8220;we also support <a href=\"https:\/\/docs.pingcap.com\/ai\/vector-search-hybrid-search\/\">hybrid search<\/a>,&#8221; even though that capability matters. The argument is infrastructure:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Mutable memory needs strong consistency.<\/strong> Cognitive memory in mem9 is revised, superseded, expired, and purged. Workspace state in drive9 advances through file revisions and background enrichment. TiDB gives that mutable state a transactional foundation, and TiKV provides the distributed, Raft-replicated storage beneath it.<\/li>\n\n\n\n<li><strong>Memory is metadata-heavy, not just vector-heavy.<\/strong> Scopes, revisions, timestamps, tags, state, policy, and provenance matter at least as much as similarity search. A SQL-native platform handles that shape more naturally than a vector-only design.<\/li>\n\n\n\n<li><strong>One platform lowers agent failure rates.<\/strong> A consolidated substrate for mem9, drive9, and operational control reduces sync errors and gives the agent a cleaner view of reality. Better memory quality often comes from fewer data boundaries, not just smarter retrieval.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/www.pingcap.com\/ko\/ai\/vector-search\/\">Hybrid retrieval<\/a> still has a place, especially for recall paths that combine exact matches, semantic search, and metadata filters. But it is a supporting capability. The primary win is that cognitive memory and workspace state live on one operational substrate. For teams building agent infrastructure, this is the same case <a href=\"https:\/\/docs.pingcap.com\/ai\/\">TiDB makes for agentic AI workloads<\/a> generally: Scale and state on a single engine instead of a stitched-together stack.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"What_Engineering_Teams_Should_Build_Next\"><\/span><strong>What Engineering Teams Should Build Next<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The near-term opportunity is not the biggest memory bucket. It is the most reliable memory operating model.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Separate cognitive memory from workspace memory instead of treating all recalled context as interchangeable.<\/li>\n\n\n\n<li>Design type, scope, and time together so memory stays useful without becoming ungovernable.<\/li>\n\n\n\n<li>Make ingestion and forgetting first-class decisions, not background cleanup problems.<\/li>\n\n\n\n<li>Keep provenance attached so facts can always trace back to artifacts, events, or human decisions.<\/li>\n\n\n\n<li>Consolidate the substrate where possible so agents spend less effort stitching reality back together.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">The industry spent the last year proving that agents can use tools. The next is about proving that agents can remember responsibly. Once memory becomes central to how agents personalize, reason, collaborate, and recover, it stops being a prompt engineering trick and becomes part of the runtime. That runtime has two sides, cognitive and workspace, and mem9 plus drive9 is our answer for building both on one operationally sound foundation.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Build agent memory on infrastructure designed for production. <a href=\"https:\/\/www.pingcap.com\/ko\/ai\/agentic-ai\/\">Learn how TiDB supports AI agent workloads<\/a> with cognitive memory, workspace state, hybrid retrieval, and governance on one durable foundation.<\/p>","protected":false},"excerpt":{"rendered":"<p>Ask a production agent a simple operational question and the cracks show immediately. What did this agent know last Tuesday? Which facts were true before a customer changed their policy? Which records must we purge everywhere because they contain sensitive data, including copies in a semantic index and a restored snapshot? Most agent stacks today [&hellip;]<\/p>\n","protected":false},"author":291,"featured_media":33704,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[145],"tags":[490,499,483,500,31],"class_list":["post-33693","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-thought-leadership","tag-agent-memory","tag-agent-state","tag-ai-agents","tag-ai-workloads","tag-tidb-cloud"],"acf":[],"featured_image_src":"https:\/\/static.pingcap.com\/files\/2026\/06\/04132423\/Blog-Feature.png","author_info":{"display_name":"Fan Wang","author_link":"https:\/\/www.pingcap.com\/ko\/blog\/author\/fwang\/"},"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.6 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>AI Agent Memory: Runtime for Production Agents<\/title>\n<meta name=\"description\" content=\"Why AI agent memory needs runtime infrastructure, not just context windows, vector stores, or retrieval add-ons.\" \/>\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\/ai-agent-memory-runtime-production-agents\/\" \/>\n<meta property=\"og:locale\" content=\"ko_KR\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"AI Agent Memory: Runtime for Production Agents\" \/>\n<meta property=\"og:description\" content=\"Why AI agent memory needs runtime infrastructure, not just context windows, vector stores, or retrieval add-ons.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.pingcap.com\/ko\/blog\/ai-agent-memory-runtime-production-agents\/\" \/>\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=\"2026-06-03T20:33:20+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2026-06-06T00:16:28+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/static.pingcap.com\/files\/2026\/06\/04132438\/Blog-LinkedIn.png\" \/>\n\t<meta property=\"og:image:width\" content=\"1200\" \/>\n\t<meta property=\"og:image:height\" content=\"627\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"author\" content=\"Fan Wang\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:image\" content=\"https:\/\/static.pingcap.com\/files\/2026\/06\/04132503\/Blog-Twitter-Banner.png\" \/>\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=\"Fan Wang\" \/>\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\\\/ai-agent-memory-runtime-production-agents\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/ai-agent-memory-runtime-production-agents\\\/\"},\"author\":{\"name\":\"Fan Wang\",\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/#\\\/schema\\\/person\\\/3e3ecea7e7a889f3a4c5ec0a67ef7561\"},\"headline\":\"Beyond Context Windows: Memory is the Runtime for Production Agents\",\"datePublished\":\"2026-06-03T20:33:20+00:00\",\"dateModified\":\"2026-06-06T00:16:28+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/ai-agent-memory-runtime-production-agents\\\/\"},\"wordCount\":1813,\"publisher\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/#organization\"},\"image\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/ai-agent-memory-runtime-production-agents\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/static.pingcap.com\\\/files\\\/2026\\\/06\\\/04132423\\\/Blog-Feature.png\",\"keywords\":[\"Agent Memory\",\"Agent State\",\"AI Agents\",\"AI Workloads\",\"TiDB Cloud\"],\"articleSection\":[\"Thought Leadership\"],\"inLanguage\":\"ko-KR\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/ai-agent-memory-runtime-production-agents\\\/\",\"url\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/ai-agent-memory-runtime-production-agents\\\/\",\"name\":\"AI Agent Memory: Runtime for Production Agents\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/ai-agent-memory-runtime-production-agents\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/ai-agent-memory-runtime-production-agents\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/static.pingcap.com\\\/files\\\/2026\\\/06\\\/04132423\\\/Blog-Feature.png\",\"datePublished\":\"2026-06-03T20:33:20+00:00\",\"dateModified\":\"2026-06-06T00:16:28+00:00\",\"description\":\"Why AI agent memory needs runtime infrastructure, not just context windows, vector stores, or retrieval add-ons.\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/ai-agent-memory-runtime-production-agents\\\/#breadcrumb\"},\"inLanguage\":\"ko-KR\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/ai-agent-memory-runtime-production-agents\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"ko-KR\",\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/ai-agent-memory-runtime-production-agents\\\/#primaryimage\",\"url\":\"https:\\\/\\\/static.pingcap.com\\\/files\\\/2026\\\/06\\\/04132423\\\/Blog-Feature.png\",\"contentUrl\":\"https:\\\/\\\/static.pingcap.com\\\/files\\\/2026\\\/06\\\/04132423\\\/Blog-Feature.png\",\"width\":1800,\"height\":600},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/www.pingcap.com\\\/blog\\\/ai-agent-memory-runtime-production-agents\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/www.pingcap.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Beyond Context Windows: Memory is the Runtime for Production Agents\"}]},{\"@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\\\/3e3ecea7e7a889f3a4c5ec0a67ef7561\",\"name\":\"Fan Wang\",\"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\":\"Fan Wang\"},\"description\":\"VP of Engineering &amp; AI Growth\",\"url\":\"https:\\\/\\\/www.pingcap.com\\\/ko\\\/blog\\\/author\\\/fwang\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"AI Agent Memory: Runtime for Production Agents","description":"Why AI agent memory needs runtime infrastructure, not just context windows, vector stores, or retrieval add-ons.","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\/ai-agent-memory-runtime-production-agents\/","og_locale":"ko_KR","og_type":"article","og_title":"AI Agent Memory: Runtime for Production Agents","og_description":"Why AI agent memory needs runtime infrastructure, not just context windows, vector stores, or retrieval add-ons.","og_url":"https:\/\/www.pingcap.com\/ko\/blog\/ai-agent-memory-runtime-production-agents\/","og_site_name":"TiDB","article_publisher":"https:\/\/facebook.com\/pingcap2015","article_published_time":"2026-06-03T20:33:20+00:00","article_modified_time":"2026-06-06T00:16:28+00:00","og_image":[{"width":1200,"height":627,"url":"https:\/\/static.pingcap.com\/files\/2026\/06\/04132438\/Blog-LinkedIn.png","type":"image\/png"}],"author":"Fan Wang","twitter_card":"summary_large_image","twitter_image":"https:\/\/static.pingcap.com\/files\/2026\/06\/04132503\/Blog-Twitter-Banner.png","twitter_creator":"@PingCAP","twitter_site":"@PingCAP","twitter_misc":{"\uae00\uc4f4\uc774":"Fan Wang","\uc608\uc0c1 \ub418\ub294 \ud310\ub3c5 \uc2dc\uac04":"9\ubd84"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.pingcap.com\/blog\/ai-agent-memory-runtime-production-agents\/#article","isPartOf":{"@id":"https:\/\/www.pingcap.com\/blog\/ai-agent-memory-runtime-production-agents\/"},"author":{"name":"Fan Wang","@id":"https:\/\/www.pingcap.com\/#\/schema\/person\/3e3ecea7e7a889f3a4c5ec0a67ef7561"},"headline":"Beyond Context Windows: Memory is the Runtime for Production Agents","datePublished":"2026-06-03T20:33:20+00:00","dateModified":"2026-06-06T00:16:28+00:00","mainEntityOfPage":{"@id":"https:\/\/www.pingcap.com\/blog\/ai-agent-memory-runtime-production-agents\/"},"wordCount":1813,"publisher":{"@id":"https:\/\/www.pingcap.com\/#organization"},"image":{"@id":"https:\/\/www.pingcap.com\/blog\/ai-agent-memory-runtime-production-agents\/#primaryimage"},"thumbnailUrl":"https:\/\/static.pingcap.com\/files\/2026\/06\/04132423\/Blog-Feature.png","keywords":["Agent Memory","Agent State","AI Agents","AI Workloads","TiDB Cloud"],"articleSection":["Thought Leadership"],"inLanguage":"ko-KR"},{"@type":"WebPage","@id":"https:\/\/www.pingcap.com\/blog\/ai-agent-memory-runtime-production-agents\/","url":"https:\/\/www.pingcap.com\/blog\/ai-agent-memory-runtime-production-agents\/","name":"AI Agent Memory: Runtime for Production Agents","isPartOf":{"@id":"https:\/\/www.pingcap.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.pingcap.com\/blog\/ai-agent-memory-runtime-production-agents\/#primaryimage"},"image":{"@id":"https:\/\/www.pingcap.com\/blog\/ai-agent-memory-runtime-production-agents\/#primaryimage"},"thumbnailUrl":"https:\/\/static.pingcap.com\/files\/2026\/06\/04132423\/Blog-Feature.png","datePublished":"2026-06-03T20:33:20+00:00","dateModified":"2026-06-06T00:16:28+00:00","description":"Why AI agent memory needs runtime infrastructure, not just context windows, vector stores, or retrieval add-ons.","breadcrumb":{"@id":"https:\/\/www.pingcap.com\/blog\/ai-agent-memory-runtime-production-agents\/#breadcrumb"},"inLanguage":"ko-KR","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.pingcap.com\/blog\/ai-agent-memory-runtime-production-agents\/"]}]},{"@type":"ImageObject","inLanguage":"ko-KR","@id":"https:\/\/www.pingcap.com\/blog\/ai-agent-memory-runtime-production-agents\/#primaryimage","url":"https:\/\/static.pingcap.com\/files\/2026\/06\/04132423\/Blog-Feature.png","contentUrl":"https:\/\/static.pingcap.com\/files\/2026\/06\/04132423\/Blog-Feature.png","width":1800,"height":600},{"@type":"BreadcrumbList","@id":"https:\/\/www.pingcap.com\/blog\/ai-agent-memory-runtime-production-agents\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.pingcap.com\/"},{"@type":"ListItem","position":2,"name":"Beyond Context Windows: Memory is the Runtime for Production Agents"}]},{"@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\/3e3ecea7e7a889f3a4c5ec0a67ef7561","name":"Fan Wang","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":"Fan Wang"},"description":"VP of Engineering &amp; AI Growth","url":"https:\/\/www.pingcap.com\/ko\/blog\/author\/fwang\/"}]}},"grav_blocks":false,"card_markup":"<a class=\"card-resource bg-white\" href=\"https:\/\/www.pingcap.com\/ko\/blog\/ai-agent-memory-runtime-production-agents\/\"><div class=\"card-resource__image-container\"><img class=\"card-resource__image\" alt=\"Blog - Feature\" src=\"https:\/\/static.pingcap.com\/files\/2026\/06\/04132423\/Blog-Feature.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\">Beyond Context Windows: Memory is the Runtime for Production Agents<\/h5><\/div><\/a>","_links":{"self":[{"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/posts\/33693","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\/291"}],"replies":[{"embeddable":true,"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/comments?post=33693"}],"version-history":[{"count":21,"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/posts\/33693\/revisions"}],"predecessor-version":[{"id":33757,"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/posts\/33693\/revisions\/33757"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/media\/33704"}],"wp:attachment":[{"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/media?parent=33693"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/categories?post=33693"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.pingcap.com\/ko\/wp-json\/wp\/v2\/tags?post=33693"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}