Google gives AI agents a memory of past work

Welcome to Superpower Daily Today we're looking at how Google is giving AI agents a persistent memory of their own failed work to drastically boost their performance Let's get into it Yeah it is a fascinating shift in how we build these systems It really is I mean think about how you or I learn to do something Like learning to parallel park Every time you bump the curb you remember it you adjust But for AI agents it's basically been this endless Groundhog Day loop They bump the curb their brain gets wiped and they just do the exact same thing again Exactly until now they've been completely stateless

Yeah and if you give an agent a massive multi step research task and it fails just telling it to try again usually means it's gonna fail the exact same way Yeah Unless you manually rewrite the prompt for it Right because historically to make a model smarter researchers had to rely on what we call parametric memory Which means actually changing the model itself right Yeah exactly you have to retrain it you update the weights and parameters the actual mathematical core of the neural network And that is well it's incredibly expensive Oh massively you're talking about huge GPU clusters and weeks of compute Right it's like sending

a baseball player all the way back to the minor leagues for a year just because they struck out once Wow yeah instead of just telling them to fix their stance for the next pitch Exactly but this new framework out of Google Research which they're calling WikiSkill it fundamentally changes that They are completely separating the model's core intelligence from its procedural memory Okay let's unpack this So they aren't touching the weights at all Not at all It's basically like a sports team watching game tape of a loss to update their playbook for next week rather than sending the players all the way back to training camp

That is the perfect analogy The physical capabilities of the players don't change but the strategy does And the numbers coming out of this paper are I mean they're wild Yeah I saw the benchmark stats So using this WikiSkill setup Google as Gemini 3 5 Flash its average score across five major agent benchmarks jumped from 49 5 to 68 1 And it wasn't just Google's own models Right they tested it on Quen 3 627b the open weight model and that went from 39 4 up to 63 3 It's a massive leap without any fine tuning So how does this actually work under the hood Like how

are they building this automated game tape So they use three distinct layers The foundation is what they call the raw layer When the agent attempts a complex task say booking a flight with really specific layovers this layer just silently records the whole execution trace Like a massive log file Exactly a detailed JSON log of every single API tool it called the websites it searched what data came back and where it eventually crashed or succeeded That's the raw evidence Right then sitting right above that you have the Wiki layer and this is where a component called the Wiki Maintainer lives The Wiki Maintainer Yeah its whole

job is to analyze those messy raw logs and basically distill them into generalized lessons Give me an example of what that looks like in practice So it might notice hey every time the agent does a broad keyword search here it hits a token limit and crashes But if it filters by a specific date first it works Oh I see So it documents that specific fix into a persistent highly structured Wiki Wait I have to pause on that Wiki Maintainer thing because that sounds like a huge bottleneck waiting to happen Oh so Well if an agent runs thousands of tasks generating thousands of logs isn't that

Wiki just going to become this incredibly bloated encyclopedia It would yeah And we know if you just shove all that into a model's context window its reasoning totally degrades Like the needle in a haystack problem That is a very real constraint which brings us to the brilliance of the third layer the skill layer Oh okay The system does not just dump the whole Wiki into the prompt A component called the skill proposer reads the Wiki and translates those insights into highly targeted procedural instructions They call them skills Oh so it only pulls what it needs Exactly When the agent faces a similar task later it

only gets the specific skill relevant to that exact moment Not the entire history of its mistakes So it writes its own targeted documentation and only queries it when it's useful But and here's my question what if it learns the wrong lesson The hallucination risk Yeah Like what if it watches the game tape sees it struck out swinging and just decides okay the rule is to never swing the bat at all Right the gating mechanism they built is exactly for that A gating mechanism Yeah it doesn't just blindly push new rules to the live agent Every proposed skill goes to a separate validation data set first

like a sandbox Okay The agent tries the new instruction in the sandbox If performance drops compared to the baseline the skill's immediately rolled back It's rejected That makes sense But what happens to that rejected idea Does the system just forget it No and this is the elegant part The experience behind that failure stays in the Wiki It remembers that it tried a certain strategy and that it failed so the agent doesn't make that same mistake twice in the future That is so smart But looking at the actual data in the paper the gains are really uneven aren't they Very uneven yeah Because on math tasks

it was insane Gemini 3 5 flash jumped from 33 to 72 6 on live math Huge jump And on spreadsheet which is complex data stuff it went from 50 5 to 76 6 But then you look at long document tasks like Office QA and the gains were tiny Why does this work so well for math but just completely flop for reading a long report Well it reveals a lot about how procedural memory works Math and spreadsheets are highly deterministic Step by step Right there are clear logical gates If the agent messes up algebra it's usually a very specific error like forgetting to isolate a variable

So the Wiki maintainer can easily spot that Exactly It spots the localized failure and writes a concrete rule Next time isolate the Y variable first But long document Q A is fuss Right because it just loses the context somewhere around page 50 Yeah and you can't easily write a procedural rule that says uh pay better attention to page 42 Procedural memory thrives on discrete logic not semantic comprehension That is a crucial takeaway for anyone building agents right now Like Wikiskill might not save your legal contract analyzer but for data engineering it's a massive unlock The other thing was model size The paper showed larger models

gained the most while smaller models actually struggled In some cases smaller models just abandoned these complex playbooks entirely Right because executing a multilayered playbook takes a huge baseline of reasoning Cognitive overload Yes If the skill says do step A check the JSON If X do B if Y do C a smaller model just loses the plot halfway through It can't hold all that conditional logic in its active memory It's like handing an NFL playbook to a middle school quarterback Exactly What about cross model transfer Like taking a skill learned by Gemini and giving it to Quinn They tested it It worked occasionally but it wasn't

reliable A playbook is deeply tied to a specific model's latent capabilities and biases So it's highly personalized Well Wikiskill shows how much software architecture matters for agents right now Next up on the agent software front OpenAI is trying to change how agents navigate the internet They've just launched a 10 day WebMCP challenge pushing websites to expose structured tools directly to AI agents This is a huge shift in the fundamental contract between a website and an agent Because right now web browsing for agents is incredibly brittle right Oh it's a nightmare Up until now if an agent wants to buy you shoes it uses vision models

to basically look at the screen like a human It guesses where the Add to Cart button is Right and if the website runs an A B test it moves the button two pixels to the left The whole workflow breaks Exactly We've been forcing machine intelligence to use human interfaces WebMCP which is the Web Model Context Protocol changes that Instead of treating the website as a visual workspace it lets developers define an explicit API right there in the browser So the website just says hey agent don't bother rendering the CSS Here's a clean function called Checkout Cart and exactly what inputs I need Yes and the

truly disruptive part is that it runs inside the web page's active JavaScript context Wait meaning it shares the human user's session Oh that is massive for session management So the agent doesn't need to try and juggle my passwords or bypass bot protection on a headless browser Right it operates locally right alongside you in the tab you're already logged into That's incredible And to kickstart this OpenAI teamed up with Chromium CloudFlare Shopify Vercel Render and Netlify for a 35K prize pool contest Closing September 3rd with winners announced the 23rd Right but here's where it gets really interesting But I have to ask are developers really gonna

rewrite their sites for this when it's just an experimental flag That is the central bottleneck Because right now only Chrome 1 46 even has a developer trial for it right Yes and it's hidden behind a manual flag Firefox Safari Edge none of them have shipped implementation So it's a total chicken and egg problem Completely Browsers won't enable it by default until major sites use it and major sites won't build for it until it's standard in Safari and Chrome Also it's very narrow in scope right now Oh so It only exposes callable tools It leaves out the broader resources and prompts that you find in the

full backend model context protocol It's very much a minimum viable product Got it So the real question for this contest is whether developers actually discover enough useful actions to justify supporting it across the wider web Exactly If Shopify realizes they can boost conversion by 15 because agents stop making visual errors at checkout you'll see a massive push for standard adoption Oh it'll be fascinating to see how fast the web adapts to that automation Meanwhile as these tools get more autonomous the policy world is trying to figure out the economic fallout and Bill Gates has just weighed in He is calling for taxes on both AI

tokens and robots while reserving specific jobs exclusively for humans Yeah in a new Gates Notes essay he essentially took his 2017 robot tax idea and gave it a major 2026 upgrade So he wants to tax the tokens themselves Yes the AI tokens the units of processing His foundational argument is about the disparity in how we tax labor versus capital Right because if I hire a human I pay payroll taxes Which funds the social safety net But if you buy a robotic arm or subscribe to an enterprise AI the tax code treats that as a business expense You write it off So he's arguing that the

system actively incentivizes companies to replace humans Exactly So he wants to use this new tax revenue to fund aggressive retraining and safety nets If we connect this to the bigger picture this is an attempt to use the tax code as a pacing mechanism for societal disruption To slow it down Right allowing a gradual phase in over decades rather than an overnight replacement of the workforce But the mechanics of this sound impossible How do you even tax a token If I'm running open source Llama 3 locally on a Mac under my desk is the IRS gonna put a meter on my graphics card That is the

most glaring vulnerability There's no governing framework no defined tax authority no clear definition of what a taxable robot even is compared to a standard software script It's a definitional nightmare And that ties into his other idea a human reserved category for jobs where AI use would be restricted or banned He used the example of a doctor telling a patient they have an incurable disease Which requires human empathy Right So what does this all mean in practice Who actually gets to decide which jobs qualify as human reserved Well that would trigger the most vicious lobbying battle in history Every professional guild would march on Washington arguing

their specific job requires the human touch Yeah the Bar Association medical boards Yeah It'd be chaos And the International Federation of Robotics is already pushing back hard aren't they Very hard They stated it would make investment artificially expensive and severely hamper national productivity and competitiveness Because if the US taxes AI heavily and China doesn't the innovation just moves overseas Exactly And historically the EU rejected a similar robot tax in 2017 South Korea has a 2018 law tweaking incentives but is not a direct tax So we'll have to watch whether this remains just a conceptual talking point or if actual policymakers attempt to drive these boundaries

Right Well that hypothetical about sensitive medical interactions perfectly sets up our next story Shifting gears from hypothetical medical policy to actual medical practice surgeons at London's National Hospital for Neurology and Neurosurgery have just removed an 11 millimeter brain tumor using a live AI anatomy mapping system This is a remarkable clinical milestone It really is So in May a 48 year old patient named Rhys Hibbert had a pituitary tumor successfully removed And the AI system was analyzing the live camera footage from inside his skull in real time Right It uses real time semantic segmentation It color codes critical anatomy the pituitary gland the nerves the major

blood vessels and exactly how the tissue is interacting with the surgical instruments It sounds exactly like a heads up display in a fighter jet It highlights all the critical targets and dangers but the pilot is absolutely still the one flying the plane That is the crucial distinction The AI made zero decisions The surgeons remained entirely in control But the stakes are just astronomical Oh absolutely The pituitary is right next to the optic nerves and major arteries A one millimeter error can mean permanent blindness a massive stroke or death Wow So having an AI that has been trained on hundreds of surgical videos way more cases

than a single surgeon would normally see in a lifetime it provides this crucial recognition layer to prevent those microscopic errors It's cognitive offloading for the surgeon Exactly And Hibbert's recovery was amazing But we have to remember this is a clinical trial milestone right It's a single data point It's not a proven population level advantage yet Right The broader clinical trial which is funded by the National Institute for Health and Care Research is what we need to watch We have to see if this live visual guidance definitively improves outcomes over standard practice across the board It just highlights the incredible real world compute required to run

live video analysis in an operating room Returning to the actual infrastructure powering all these models OpenAI has revealed its first custom inference chip called Jalapeño claiming it drastically outperforms comparison systems on efficiency This was presented at the Hot Chips Conference and it is a major pivot toward vertical integration for them So Jalapeño delivered 1 5 to 1 9 times more AI work per watt and 1 7 to 3 6 times lower end to end latency compared to other systems And on one model Kimi K2 51T they actually hit 3 4 times lower latency That is fast And it hares a 700 watt chip that stayed

at or below 550 watts during testing How did they pull this off They co designed the entire stack the chip memory networking and software and the key was keeping the model state and the KV cache completely local on the ship Okay remind me why the KV cache is the big bottleneck Well inference has two phases First the compute heavy pre fill phase where it processes your prompt then the memory bandwidth heavy decode phase where it generates the answer one token at a time If the chip has to constantly fetch data back and forth from external memory for that decode phase you get a massive latency

bottleneck The KV cache is the model's short term memory of the conversation So by keeping that local they eliminated all that data handoff Exactly And for interactive agents doing 10 sequential tasks a 3x speedup on each step is the difference between a usable product and a totally broken user experience Okay I have to play devil's advocate here though It's easy to win on your own cherry picked benchmarks Is this actually gonna replace NVIDIA hardware for them Probably not replace no Because these tests were highly controlled right They used the nominal 8K prompts and 1K responses in single turn workloads They completely excluded long context multi

turn Agent X scenarios that would actually stress that local cache routing Yes those massive multi turn workflows would likely spill out of the local cache and slow everything down Plus it's still just an engineering sample Richard Ho OpenAI's head of hardware said scale won't happen until 2027 Right and they're gonna continue deploying NVIDIA heavily So what we need to watch is how this actually holds up in 2027 production especially since NVIDIA's Vera Rubin systems will be hitting the market by then The silicon wars are only gonna get hotter next year Definitely Let's round out the episode with some rapid fire updates across the ecosystem We'll

hit MIT the Cyber Coalition the Beijing Games and Cambridge We've got breakthroughs in proteins cyber threats robots and pure math Let's start with biology MIT researchers Foster Birnbaum and Amy Keating built a system called POTS MPNN And this changes how AI designs proteins right Fundamentally yeah Instead of just copying natural evolutionary sequences like what nature has already done it evaluates all 20 amino acid options across position pairs to see if a sequence actually fits a target structure's energy landscape Oh so it prioritizes physics and stability over just evolutionary patterns Exactly though the caveat is it still relies somewhat on native sequence training data to get

started Still it's a huge leap Okay next up the Cyber Defense Coalition OpenAI Anthropic CrowdStrike and over 100 other firms just issued a massive warning about imminent AI cyber attacks They are urging governments to provide defensive AI to water utilities and hospitals And the timing here is no coincidence because CISA just reported that over 100 US water systems were hit in July The attackers used AI assisted scripts to target internet exposed PLCs programmable logic controllers Which physically control the water valves and chemicals Right but the caveat here is that this letter is just a warning There's no funding attached no deployment plans just an alarm

bell Which leads us to the Beijing Humanoid Games where the physical reality of robots was put to the test And it was a brutal test Out of 12 teams only three finished a 30 minute fire and rescue course involving shutting valves and using an extinguisher But the household test was even more telling Yeah the unplanned package delivery The robots were doing laundry someone dropped off a package and it completely disrupted their logic It just shows the massive gap between polished athletic demos like X Humanoid winning a 400 meter race through simulation learning and actually handling interrupted messy real world chores Totally different ballgame Finally let's

talk math Cambridge mathematician Henry Bradford claims OpenAI's ASTRA model has solved the non celetic group's existence problem This is a highly complex open problem in pure math But the nuance is that ASTRA didn't invent a totally new framework right Correct It found an incredibly clever twist on existing theorems by mathematicians Kuhn and Thom Still it raises huge existential questions for universities Bradford brought up that old essay by William Thurston about the nature of math If AI can pump out cheap perfectly accurate theorems will universities still value human mathematicians Or just treat them as expendable output generators It's a heavy question It really is All right

that brings us to the end Here are your three concise takeaways Number one First software design is yielding massive gains independent of scaling as seen with Google's Wiki skill using procedural memory to boost agents and OpenAI creating a new API standard for the web with WebMCP Number two Second the physical integration of AI is accelerating rapidly but faces harsh realities Surgeons are successfully using it as a heads up display while humanoid robots are still struggling with basic interruptions like a package delivery Number three Third infrastructure and security are the immediate bottlenecks OpenAI is pushing custom silicon with jalapeno to reduce compounding latency while a massive

coalition is begging governments to protect vulnerable infrastructure like water systems from AI assisted scripts As for one development to watch tomorrow keep an eye on the developer response to OpenAI's WebMCP challenge The deadline is fast approaching on September 3rd and the volume of submissions will tell us if the web is actually ready to rebuild itself for agents You can find every story we discussed today and a whole lot more over at superpoweredaily com Thank you for listening and we'll see you tomorrow