--- title: Memory Harnesses for Long-Running Research Agents —... description: A Sakana AI research scientist presents experiments on building memory harnesses for long-running AI research agents running on local, on-device models. Using... canonical: https://daily.dev/posts/memory-harnesses-for-long-running-research-agents-stefania-druga-sakana-ai-kvigkw2o6 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Memory Harnesses for Long-Running Research Agents — Stefania Druga, Sakana.ai | daily.dev og:description: A Sakana AI research scientist presents experiments on building memory harnesses for long-running AI research agents running on local, on-device models. Using... og:url: https://daily.dev/posts/memory-harnesses-for-long-running-research-agents-stefania-druga-sakana-ai-kvigkw2o6 og:image: https://api.daily.dev/og/posts/KvIGkw2o6.png og:image:alt: Memory Harnesses for Long-Running Research Agents — Stefania Druga, Sakana.ai 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. # Memory Harnesses for Long-Running Research Agents — Stefania Druga, Sakana.ai **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 13 min read · 1 upvotes · 0 comments ## Summary A Sakana AI research scientist presents experiments on building memory harnesses for long-running AI research agents running on local, on-device models. Using a Mac Studio M3 Ultra running Qwen 27B (4-bit quantized) and DeepSeek V4 Flash, she tests a 'write-manage-read' memory loop with different recall strategies: no memory, vector RAG, a decisions ledger, and an oracle ground-truth baseline. On tasks that fit within the context window, memory adds cost without benefit, but on long-horizon tasks (tested on X-Bench, spanning hundreds of steps) a ranked-ledger recall policy outperforms other approaches, retrieves correct answers more often, and reduces token cost. She argues recall policy should be treated as a first-class metric in agent design and highlights growing interest in local models for cost-efficient, sovereign AI infrastructure. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What recall strategy works best for giving a long-running AI agent memory across sessions? A ranked decisions ledger, which tracks and prioritizes decisions made at every turn, outperformed no memory, vector-based RAG similarity retrieval, and simple memory-gating in experiments on the X-Bench long-horizon benchmark across 68 questions and multiple seeds. It beat gating (deciding whether to use memory at all) and also reduced token cost compared to unranked or absent memory. _Engineers designing agent memory systems can track emerging recall-policy techniques like this one on daily.dev._ ### Does adding a memory system always improve an AI agent's task performance? No, memory only helps when the relevant information does not fit within the model's context window. In literature-review tasks where all papers fit in context, adding a memory harness gave identical performance to no memory at all, while increasing cost. Memory harnesses pay off specifically on long-horizon tasks where needed information falls outside the context window. _Anyone deciding whether to add memory infrastructure to an agent can weigh trade-offs like these via daily.dev._ ### What local hardware setup can run models like Qwen 27B and DeepSeek V4 Flash for agent experiments? An Apple M3 Ultra Mac with 96GB RAM and 28 core CPUs was used to run Qwen 27B quantized at 4-bit alongside DeepSeek V4 Flash for long-horizon agent memory experiments. A known constraint is that DeepSeek V4 Flash does not support batch querying, so evaluations must run serially, making multi-day runs necessary even with capable hardware. _Developers evaluating local model setups for agentic workloads can follow practical hardware notes like this on daily.dev._ ## Similar posts on daily.dev - [Building MindCache: Designing an Agentic Memory System for Long-Term AI](https://daily.dev/posts/building-mindcache-designing-an-agentic-memory-system-for-long-term-ai-k9xwjuamo) · Medium · 1 upvotes · 0 comments - [Agent Memory: Why Your AI Has Amnesia and How to Fix It](https://daily.dev/posts/agent-memory-why-your-ai-has-amnesia-and-how-to-fix-it-nkpwfuhgz) · Tech Lead Digest · 1 upvotes · 0 comments - [How to improve the memory of AI agents](https://daily.dev/posts/how-to-improve-the-memory-of-ai-agents-zmppe2bgj) · InfoWorld · 0 upvotes · 0 comments - [Retrieval vs. Memory in Agentic AI System](https://daily.dev/posts/retrieval-vs-memory-in-agentic-ai-system-w27x4mch5) · Machine Learning Mastery · 2 upvotes · 0 comments - [Memory for AI Agents: A New Paradigm of Context Engineering](https://daily.dev/posts/memory-for-ai-agents-a-new-paradigm-of-context-engineering-ku2rvmlcw) · The New Stack · 2 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#rag](https://daily.dev/tags/rag), [#agentic-ai](https://daily.dev/tags/agentic-ai) [View this post on daily.dev](https://daily.dev/posts/memory-harnesses-for-long-running-research-agents-stefania-druga-sakana-ai-kvigkw2o6) ```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":"Memory Harnesses for Long-Running Research Agents — Stefania Druga, Sakana.ai","url":"https://daily.dev/posts/memory-harnesses-for-long-running-research-agents-stefania-druga-sakana-ai-kvigkw2o6","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/memory-harnesses-for-long-running-research-agents-stefania-druga-sakana-ai-kvigkw2o6"},"datePublished":"2026-08-12T15:24:11.709Z","dateModified":"2026-09-14T06:05:00.615Z","description":"A Sakana AI research scientist presents experiments on building memory harnesses for long-running AI research agents running on local, on-device models. Using...","image":"https://i.ytimg.com/vi/R3-anFK1YM8/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/R3-anFK1YM8/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/memory-harnesses-for-long-running-research-agents-stefania-druga-sakana-ai-kvigkw2o6","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,rag,agentic-ai","timeRequired":"PT13M","video":{"@type":"VideoObject","name":"Memory Harnesses for Long-Running Research Agents — Stefania Druga, Sakana.ai","description":"A Sakana AI research scientist presents experiments on building memory harnesses for long-running AI research agents running on local, on-device models. Using...","thumbnailUrl":"https://i.ytimg.com/vi/R3-anFK1YM8/sddefault.jpg","uploadDate":"2026-08-12T15:24:11.709Z","duration":"PT13M","url":"https://api.daily.dev/r/KvIGkw2o6","embedUrl":"https://www.youtube.com/embed/R3-anFK1YM8"}} {"@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":"Memory Harnesses for Long-Running Research Agents — Stefania Druga, Sakana.ai"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/memory-harnesses-for-long-running-research-agents-stefania-druga-sakana-ai-kvigkw2o6#faq","mainEntity":[{"@type":"Question","name":"What recall strategy works best for giving a long-running AI agent memory across sessions?","acceptedAnswer":{"@type":"Answer","text":"A ranked decisions ledger, which tracks and prioritizes decisions made at every turn, outperformed no memory, vector-based RAG similarity retrieval, and simple memory-gating in experiments on the X-Bench long-horizon benchmark across 68 questions and multiple seeds. It beat gating (deciding whether to use memory at all) and also reduced token cost compared to unranked or absent memory. Engineers designing agent memory systems can track emerging recall-policy techniques like this one on daily.dev."}},{"@type":"Question","name":"Does adding a memory system always improve an AI agent's task performance?","acceptedAnswer":{"@type":"Answer","text":"No, memory only helps when the relevant information does not fit within the model's context window. In literature-review tasks where all papers fit in context, adding a memory harness gave identical performance to no memory at all, while increasing cost. Memory harnesses pay off specifically on long-horizon tasks where needed information falls outside the context window. Anyone deciding whether to add memory infrastructure to an agent can weigh trade-offs like these via daily.dev."}},{"@type":"Question","name":"What local hardware setup can run models like Qwen 27B and DeepSeek V4 Flash for agent experiments?","acceptedAnswer":{"@type":"Answer","text":"An Apple M3 Ultra Mac with 96GB RAM and 28 core CPUs was used to run Qwen 27B quantized at 4-bit alongside DeepSeek V4 Flash for long-horizon agent memory experiments. A known constraint is that DeepSeek V4 Flash does not support batch querying, so evaluations must run serially, making multi-day runs necessary even with capable hardware. Developers evaluating local model setups for agentic workloads can follow practical hardware notes like this on daily.dev."}}]} ```