{"slug": "i-dont-think-im-losing-my-memory-i-think-im-trying-to-think-at-ai-speed", "title": "I Don’t Think I’m Losing My Memory. I Think I’m Trying to Think at AI Speed.", "summary": "A developer reports that using multiple AI coding agents in parallel is causing them to frequently forget small tasks and lose context, a phenomenon they attribute not to memory loss but to the human brain's limited working memory being overwhelmed by AI's ability to parallelize work. The engineer notes that while AI reduces the cost of producing work, it does not reduce the cognitive cost of understanding, evaluating, and integrating results, leading to a feeling of producing more while being less focused.", "body_md": "Something strange has been happening lately, I keep forgetting small things.\n\nNot names.\n\nNot conversations.\n\nNot important events.\n\nSmall pieces of routines.\n\nThings I’ve done almost automatically for years, for example:\n\nEvery night I usually prepare everything I need for the gym the next morning.\n\nClothes.\n\nShoes.\n\nSupplements.\n\nNothing complicated, but lately I’ll prepare the clothes... and forget the shoes. Or prepare everything except the supplements.\n\nSmall things, but frequent enough that I started noticing. And then I started seeing the same pattern at work.\n\nI might have three or four coding agents running at the same time.\n\nOne is fixing a bug.\n\nAnother is exploring a ticket.\n\nAnother is refactoring something.\n\nAnother is working in a different application entirely.\n\nWhile Agent A works, I move to B, while B is thinking, I review C, then A finishes, so I go back to A.\n\nAnd suddenly:\n\n**What was I doing before this?**\n\nSometimes I need a minute, I open the previous ticket, read a few lines. look at the code and then the entire context comes back.\n\nThat’s when I started wondering:\n\n**Maybe this isn’t a memory problem.**\n\nMaybe I’m just constantly dropping context.\n\nWe talk about context all the time. processes have context, requests have context, applications have state, LLMs have context windows.\n\nAnd humans do too.\n\nWhen I’m working on a bug, my brain is temporarily holding things like:\n\nThat information may never reach long-term memory, It doesn’t need to. It only needs to stay available long enough for me to finish the task.\n\nThat is working memory and working memory is not infinite.\n\nFor most of my career, software development had friction.\n\nYou wrote code, you compiled, you waited, you ran tests, you searched documentation, you read Stack Overflow, you tried something. It failed, you stared at the problem, then you tried again, there were pauses. Not necessarily intentional pauses, but pauses.\n\nNow I can do this:\n\n``` php\nTask A -> Agent A working\nTask B -> Agent B working\nTask C -> Agent C working\nTask D -> Agent D working\n```\n\nAmazing, four things moving at once, except eventually:\n\n``` php\nAgent A -> finished\nAgent B -> needs clarification\nAgent C -> changed 12 files\nAgent D -> found another problem\n```\n\nAnd every single one of them comes back to the same place.\n\nMe.\n\nThis sounds obvious, but I don’t think I fully understood what it meant. AI made execution parallel, it did not make my attention parallel.\n\nEvery result still needs a human, someone has to understand it, review it, question it, check the diff, understand the architecture, notice the weird assumption, decide what happens next.\n\nAI reduced the cost of producing work, but it did not remove the cost of:\n\n**understanding**\n\n**evaluating**\n\n**integrating**\n\n**deciding**\n\nAnd now there is simply more of that work arriving.\n\nFaster.\n\nThis is where things get strange. Imagine a task used to take me 30 minutes. Now an agent gives me a decent first implementation in 3, I saved 27 minutes.\n\nGreat.\n\nSo what do I do with those 27 minutes?\n\nI start another task of course - duh!\n\nThen another, then another and suddenly my workflow looks like this:\n\n```\nprompt\nswitch\nreview\nswitch\napprove\nswitch\ndebug\nswitch\nask follow-up\nswitch\nread diff\nswitch\nremember what I was doing\n```\n\nThe machine got faster, so I made my day more concurrent, and eventually something interesting happens:\n\n**I am producing more while feeling less focused.**\n\nThat feels like a contradiction, it probably isn’t.\n\nWe already knew this before AI.\n\nSlack did it.\n\nEmail did it.\n\nMeetings did it.\n\nNotifications did it.\n\nBut agents make something much easier:\n\n**starting more work.**\n\nBefore AI, opening another task meant I had to actually work on another task, there was a natural cost. Now I can start another workstream with a prompt, thirty seconds later, something else is running in parallel.\n\nThe cost of starting work has collapsed, the cost of keeping that work inside my head has not. That changes everything.\n\nThis is probably the part I find most interesting. When I return to something after ten minutes and think:\n\nWhat was I doing?\n\nthe experience feels like forgetting, but then I look at the code.\n\nRead the ticket.\n\nSee the previous message.\n\nAnd suddenly:\n\nOh, right.\n\nEverything comes back, the memory was probably not gone. The context was simply no longer active.\n\nThere’s a big difference between:\n\n**losing information** and **losing access to information because your attention moved somewhere else.**\n\nSubjectively, though, they can feel almost identical.\n\nBoth feel like:\n\nI forgot.\n\nI don’t think AI explains everything.\n\nNot even close, stress matters, sleep matters, family matters. mental health matters, changes in routine matter.\n\nLife can leave a lot of invisible processes running in the background.\n\nIn my own case, the last few months have contained enough personal stuff that cognitive load is probably already higher than normal.\n\nSo saying:\n\nAI is doing this to me\n\nwould be lazy. I don’t know that, but I do think AI may be acting as an amplifier. It gives me the ability to create far more cognitive demand than I could create before.\n\nAnd because that demand looks like productive work... it’s very easy not to recognize it as overload.\n\nThat’s probably what worries me most, overload doesn’t always feel like overload.\n\nSometimes it looks like this:\n\nEverything appears productive, meanwhile:\n\nattention becomes more fragmented, tasks become harder to reconstruct, small routines start losing pieces, you start rereading things, you switch windows and forget why, you finish the day having done a lot... but feeling strangely scattered.\n\nThat’s an interesting failure mode because the dashboard says productivity went up.\n\nFor decades, engineering has tried to remove latency.\n\nFaster builds.\n\nFaster tests.\n\nFaster computers.\n\nFaster deployments.\n\nFaster feedback.\n\nFaster everything.\n\nThat was mostly a good thing, but I’ve started wondering whether some of that waiting had an accidental benefit.\n\nWaiting for a build.\n\nWalking to get coffee.\n\nReading documentation slowly.\n\nLooking at a bug without immediately asking something else for the answer.\n\nSitting with a problem. Those moments looked inefficient, but perhaps they gave the brain time to stay inside one context.\n\nAI is removing enormous amounts of friction, which is amazing. But the machine no longer forces us to slow down, maybe that means we need to do it ourselves.\n\nThis is the adjustment I’m experimenting with.\n\nI’m not trying to use less AI. I’m not going back to doing everything manually, that would miss the point.\n\nInstead, I’m trying to separate two resources:\n\n```\nMachine throughput\n!=\nHuman attention\n```\n\nAn agent finishing something does not mean I must review it immediately. An agent waiting does not mean I must start another one. A task taking 3 minutes instead of 30 does not mean I need 10 tasks in flight.\n\nThe mental model I’m trying to adopt is:\n\n```\nAI works in parallel.\n\nI work sequentially.\n```\n\nThe agents can wait.\n\nThe output can wait.\n\nThe notification can wait.\n\nMy attention decides what runs next.\n\nThe funny thing is that most agent interfaces still behave like chat applications.\n\nSomething happens, they tell you. Another thing happens, they tell you again.\n\nThe result is basically:\n\n``` php\nAgent A -> INTERRUPT\nAgent C -> INTERRUPT\nAgent B -> INTERRUPT\nAgent D -> INTERRUPT\n```\n\nMaybe that is the wrong model, maybe what I actually want is:\n\n```\n11:00 AM\n\nAgent review queue\n\nA: finished\nB: blocked\nC: needs one decision\nD: still working\n```\n\nOne review window, one context-switching event. Machines stay asynchronous and humans stay sane. That feels much closer to how agent systems should eventually work.\n\nI suspect developers will encounter this problem earlier than most people, simply because coding agents already make extreme parallelism possible.\n\nToday I can have agents:\n\nall at once.\n\nSoon this won’t be unusual, and it won’t be limited to developers.\n\nManagers will have agents, designers will have agents, researchers will have agents, founders will have agents, students will have agents. Knowledge work may become massively parallel, human cognition will not.\n\nThere is already a lot of discussion around cognitive offloading.\n\nThe concern is understandable.\n\nIf AI writes for us, remembers for us, searches for us, reasons for us... maybe we exercise certain cognitive abilities less, but there may be another problem hiding on the opposite side.\n\nNot:\n\n**AI makes us think less.**\n\nBut:\n\n**AI makes it possible for us to try thinking about too many things at once.**\n\nThe chain might look something like this:\n\n```\nMore AI\n   ↓\nMore output\n   ↓\nMore work in flight\n   ↓\nMore decisions\n   ↓\nMore context switching\n   ↓\nMore fragmented attention\n```\n\nMaybe we’re not becoming forgetful, maybe we’re just overflowing our cognitive context window.\n\nThis may sound obvious too, but I think many of us are unconsciously doing exactly that.\n\nThe AI answers faster, so we answer faster. The agent finishes faster, so we start something else faster. Research takes seconds, so we immediately consume more research.\n\nEvery efficiency gain creates another opportunity to do something.\n\nAnd if every second AI saves becomes another task... we haven’t really created free time, we’ve increased throughput. Those are not the same thing.\n\nIt’s still early, agents are new and most people are only starting to build AI deeply into their daily workflows. So I don’t know whether this becomes a serious widespread proble, maybe it doesn’t, but I think it’s worth paying attention to, because eventually one developer won’t be supervising one assistant.\n\nThey may be supervising five agents, then ten, then twenty. At some point, the machine will not be the limiting factor anymore. The limiting factor will be the person trying to understand what all those machines just did.\n\nAnd maybe one of the most important skills in an AI-heavy world will not be learning how to make AI move faster.\n\nIt will be learning when **not to follow it**.\n\nThe machines can work at machine speed, we still need to think at human speed.\n\nAnd maybe that is something worth protecting.", "url": "https://wpnews.pro/news/i-dont-think-im-losing-my-memory-i-think-im-trying-to-think-at-ai-speed", "canonical_source": "https://dev.to/jackfiallos/i-dont-think-im-losing-my-memory-i-think-im-trying-to-think-at-ai-speed-1pde", "published_at": "2026-09-02 11:10:27+00:00", "updated_at": "2026-09-02 11:25:24.841643+00:00", "lang": "en", "topics": ["ai-agents", "developer-tools", "artificial-intelligence"], "entities": [], "alternates": {"html": "https://wpnews.pro/news/i-dont-think-im-losing-my-memory-i-think-im-trying-to-think-at-ai-speed", "markdown": "https://wpnews.pro/news/i-dont-think-im-losing-my-memory-i-think-im-trying-to-think-at-ai-speed.md", "text": "https://wpnews.pro/news/i-dont-think-im-losing-my-memory-i-think-im-trying-to-think-at-ai-speed.txt", "jsonld": "https://wpnews.pro/news/i-dont-think-im-losing-my-memory-i-think-im-trying-to-think-at-ai-speed.jsonld"}}