{"slug": "what-happens-if-you-give-a-fruit-fly-brain-external-memory", "title": "What Happens If You Give a Fruit Fly Brain External Memory?", "summary": "A developer built a small recurrent neural network paired with an external memory system called Mycelium, testing whether memory can modulate neural state rather than being explicitly queried. The experiment, inspired by Google Research's connectome mapping of the fruit fly nervous system (roughly 166,700 neurons and 125 million synaptic connections) and the DOOMFLY project, enforces one rule: memory can change neural state but can never choose an action. The developer ran brain resets and memory-swap experiments between identical systems, framing the work around measurable \"individuality\" rather than consciousness.", "body_md": "Google Research and its collaborators mapped the male fruit fly nervous system: roughly 166,700 neurons and 125 million synaptic connections.\n\nThen somebody wired the thing to Doom.\n\nThe project is called DOOMFLY.\n\nNot a living fly. Not consciousness in a jar. The biological wiring diagram is real; the dynamics and interfaces are modeled in software.\n\nCool experiment.\n\nBut it made me wonder about something else.\n\n**What happens if you give the simulated brain an external memory that it never has to explicitly query?**\n\nBecause that is the part of current AI memory systems that keeps bothering me.\n\nWe keep building memory like a tool:\n\n```\nsomething happens\nsave it\n\nsomething similar happens later\nsearch for it\n```\n\nThat works for LLM agents.\n\nBut it does not feel much like memory.\n\nI do not decide to remember.\n\nA smell drops me somewhere from 20 years ago. I walk into a room and it feels familiar before I know why. Nothing in my head calls a search function.\n\nAnd a fly definitely isn't reading a SKILL.md file telling it when to query its memory.\n\nStrip the language away and most agents look like this:\n\n```\nLLM\n+\nsystem prompt\n+\ntool definitions\n+\nSKILL.md\n+\nworkflow\n+\nmemory.search()\n+\nmemory.save()\n```\n\nThe fly has none of that. It cannot read a memory. It cannot parse JSON.\n\n```\n{\n  \"event\": \"enemy appeared\",\n  \"action\": \"turned right\",\n  \"outcome\": \"survived\"\n}\n```\n\nIf an external memory is going to change what the fly does, it has to become part of the fly's computational state. That is a harder problem. It is also a more honest one.\n\nHere is the loop the fly already runs:\n\n```\nsensory input\n     |\n     v\ncurrent neural state\n     |\n     v\nconnectome dynamics\n     |\n     v\nmotor activity\n     |\n     v\nenvironment\n```\n\nSo put the memory beside it, and never let the fly call it.\n\nThe memory system observes the state. When a similar state comes around again, it does not return a result. It feeds back in as modulation: some populations get easier to fire, some get harder, and the network resolves the rest.\n\nNo query. No lookup. The memory just becomes relevant.\n\n**This shifts the problem from retrieval to resonance.**\n\nI am not going to call this consciousness. Nobody understands consciousness well enough to claim a graph database bolted to a connectome creates it.\n\nI am going to call it individuality, because individuality is the part you can actually measure. You can put a number on how much of a behavioral identity survives a brain reset.\n\nThat is what this series is about.\n\nSo I built a tiny version.\n\nNot the full fruit fly connectome. A small recurrent network where I could control every assumption.\n\nThen I put Mycelium beside it.\n\nOne rule: **memory can change neural state. It can never choose an action.**\n\nIf the memory system says \"turn right,\" the experiment is bullshit. I have just built a bot with a fake brain attached.\n\nThe memory has to do something more subtle: make certain states easier or harder to reach, then let the network produce the behavior itself.\n\nI ran it.\n\nThen I reset the brain.\n\nThen I moved the memory into a fresh one.\n\nThen I swapped memories between two otherwise identical systems.\n\nThat last experiment is where this stopped being a fun memory demo for me.\n\nPart 2 is the code, the controls, and the numbers.\n\nClone it and tell me where I'm wrong.\n\n**Links**", "url": "https://wpnews.pro/news/what-happens-if-you-give-a-fruit-fly-brain-external-memory", "canonical_source": "https://dev.to/constant_itis/what-happens-if-you-give-a-fruit-fly-brain-external-memory-1gcm", "published_at": "2026-09-15 20:54:27+00:00", "updated_at": "2026-09-15 21:38:33.048333+00:00", "lang": "en", "topics": ["artificial-intelligence", "neural-networks", "ai-research", "ai-agents"], "entities": ["Google Research", "DOOMFLY", "Mycelium"], "alternates": {"html": "https://wpnews.pro/news/what-happens-if-you-give-a-fruit-fly-brain-external-memory", "markdown": "https://wpnews.pro/news/what-happens-if-you-give-a-fruit-fly-brain-external-memory.md", "text": "https://wpnews.pro/news/what-happens-if-you-give-a-fruit-fly-brain-external-memory.txt", "jsonld": "https://wpnews.pro/news/what-happens-if-you-give-a-fruit-fly-brain-external-memory.jsonld"}}