--- title: CrabRAG: Why Automated Assistants Need Graph Memory, Not... description: A conference talk by Stephen Chin from Neo4j arguing that AI agents relying on markdown files or vector databases for memory face fundamental limitations at... canonical: https://daily.dev/posts/crabrag-why-automated-assistants-need-graph-memory-not-more-tokens-stephen-chin-neo4j-4ycnjsvcs twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: CrabRAG: Why Automated Assistants Need Graph Memory, Not More Tokens — Stephen Chin, Neo4j | daily.dev og:description: A conference talk by Stephen Chin from Neo4j arguing that AI agents relying on markdown files or vector databases for memory face fundamental limitations at... og:url: https://daily.dev/posts/crabrag-why-automated-assistants-need-graph-memory-not-more-tokens-stephen-chin-neo4j-4ycnjsvcs og:image: https://api.daily.dev/og/posts/4YcNjSvCS.png og:image:alt: CrabRAG: Why Automated Assistants Need Graph Memory, Not More Tokens — Stephen Chin, Neo4j og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # CrabRAG: Why Automated Assistants Need Graph Memory, Not More Tokens — Stephen Chin, Neo4j **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 20 min read · 0 upvotes · 0 comments ## Summary A conference talk by Stephen Chin from Neo4j arguing that AI agents relying on markdown files or vector databases for memory face fundamental limitations at scale. Using a 'CrabRAG' demo, he compares vector-based memory against graph-based memory (backed by Neo4j via Cognee) on a home lab digital twin. The graph approach enables multi-hop reasoning and precise relationship traversal, surfacing actionable results like outdated software and exposed management ports that the vector store missed. The talk advocates for graph memory as the right architecture for autonomous agents dealing with complex, large-scale data, and promotes the upcoming 'Graph RAG: The Definitive Guide' book and Neo4j Graph Academy resources. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Why do AI agents using markdown files for memory perform poorly at scale? Markdown-file memory systems load large amounts of repetitive context on every turn just in case something is useful, often consuming over 100,000 tokens per round. This works at small scale with high-quality models but breaks down at larger scale because agents can't efficiently recall specific, relevant facts and instead reload everything indiscriminately. _daily.dev surfaces practical engineering writeups for teams weighing memory architectures in agent systems._ ### Why do vector databases cause AI agents to give wrong or hallucinated answers? Vector similarity search finds items that are related in embedding space but not necessarily factually identical or correctly linked, which leads agents to retrieve plausible-sounding but incorrect information. This problem compounds in complex, multihop reasoning scenarios where similarity alone can't trace an accurate chain of relationships, unlike graph traversal which follows explicit connections. _developers picking between vector search and graph retrieval can track real-world comparisons like this on daily.dev._ ### How does combining vector search with graph traversal improve AI agent memory accuracy? Vector search is used first to find seed nodes, then graph traversal expands from those nodes to nearest neighbors and ranks results by relatedness, producing precise multihop answers. In a home-lab security audit test using Cognee with a Neo4j backend, this approach correctly identified exposed end-of-life software and open management ports, while a pure vector-store agent returned vague, unhelpful responses on the same data. _engineers building agent memory systems can follow implementation details like this through daily.dev._ ## Similar posts on daily.dev - [Vector Databases vs. Graph RAG for Agent Memory: When to Use Which](https://daily.dev/posts/vector-databases-vs-graph-rag-for-agent-memory-when-to-use-which-8nmipkra5) · Machine Learning Mastery · 1 upvotes · 0 comments - [Knowledge graph RAG: structured retrieval for AI agents](https://daily.dev/posts/knowledge-graph-rag-structured-retrieval-for-ai-agents-k6cymxiik) · Redis · 1 upvotes · 0 comments - [GraphRAG: How AI Answers Questions Hidden Across Many Documents](https://daily.dev/posts/graphrag-how-ai-answers-questions-hidden-across-many-documents-y6prjknjz) · ByteByteGo · 7 upvotes · 1 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#rag](https://daily.dev/tags/rag), [#vector-search](https://daily.dev/tags/vector-search), [#neo4j](https://daily.dev/tags/neo4j) [View this post on daily.dev](https://daily.dev/posts/crabrag-why-automated-assistants-need-graph-memory-not-more-tokens-stephen-chin-neo4j-4ycnjsvcs) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"CrabRAG: Why Automated Assistants Need Graph Memory, Not More Tokens — Stephen Chin, Neo4j","url":"https://daily.dev/posts/crabrag-why-automated-assistants-need-graph-memory-not-more-tokens-stephen-chin-neo4j-4ycnjsvcs","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/crabrag-why-automated-assistants-need-graph-memory-not-more-tokens-stephen-chin-neo4j-4ycnjsvcs"},"datePublished":"2026-07-22T18:50:33.168Z","dateModified":"2026-09-14T07:50:50.172Z","description":"A conference talk by Stephen Chin from Neo4j arguing that AI agents relying on markdown files or vector databases for memory face fundamental limitations at...","image":"https://i.ytimg.com/vi/Q0VkgCyNVUg/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/Q0VkgCyNVUg/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/crabrag-why-automated-assistants-need-graph-memory-not-more-tokens-stephen-chin-neo4j-4ycnjsvcs","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,rag,vector-search,neo4j","timeRequired":"PT20M","video":{"@type":"VideoObject","name":"CrabRAG: Why Automated Assistants Need Graph Memory, Not More Tokens — Stephen Chin, Neo4j","description":"A conference talk by Stephen Chin from Neo4j arguing that AI agents relying on markdown files or vector databases for memory face fundamental limitations at...","thumbnailUrl":"https://i.ytimg.com/vi/Q0VkgCyNVUg/sddefault.jpg","uploadDate":"2026-07-22T18:50:33.168Z","duration":"PT20M","url":"https://api.daily.dev/r/4YcNjSvCS","embedUrl":"https://www.youtube.com/embed/Q0VkgCyNVUg"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"CrabRAG: Why Automated Assistants Need Graph Memory, Not More Tokens — Stephen Chin, Neo4j"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/crabrag-why-automated-assistants-need-graph-memory-not-more-tokens-stephen-chin-neo4j-4ycnjsvcs#faq","mainEntity":[{"@type":"Question","name":"Why do AI agents using markdown files for memory perform poorly at scale?","acceptedAnswer":{"@type":"Answer","text":"Markdown-file memory systems load large amounts of repetitive context on every turn just in case something is useful, often consuming over 100,000 tokens per round. This works at small scale with high-quality models but breaks down at larger scale because agents can't efficiently recall specific, relevant facts and instead reload everything indiscriminately. daily.dev surfaces practical engineering writeups for teams weighing memory architectures in agent systems."}},{"@type":"Question","name":"Why do vector databases cause AI agents to give wrong or hallucinated answers?","acceptedAnswer":{"@type":"Answer","text":"Vector similarity search finds items that are related in embedding space but not necessarily factually identical or correctly linked, which leads agents to retrieve plausible-sounding but incorrect information. This problem compounds in complex, multihop reasoning scenarios where similarity alone can't trace an accurate chain of relationships, unlike graph traversal which follows explicit connections. developers picking between vector search and graph retrieval can track real-world comparisons like this on daily.dev."}},{"@type":"Question","name":"How does combining vector search with graph traversal improve AI agent memory accuracy?","acceptedAnswer":{"@type":"Answer","text":"Vector search is used first to find seed nodes, then graph traversal expands from those nodes to nearest neighbors and ranks results by relatedness, producing precise multihop answers. In a home-lab security audit test using Cognee with a Neo4j backend, this approach correctly identified exposed end-of-life software and open management ports, while a pure vector-store agent returned vague, unhelpful responses on the same data. engineers building agent memory systems can follow implementation details like this through daily.dev."}}]} ```