All Talk, No Task: How Far Is an Agent From Being a Participant?

For the last couple of years, whenever Web3 talks about AI, the conversation tends to land on wallets and payments. Boiled down, the thinking goes that once an agent has a wallet it can pay, and once it can pay it can take on work and get things done.

It makes sense. Finance and settlement are what Web3 does best, and when an industry runs into a new problem, it reaches first for whatever tool sits closest to hand.

There has been no shortage of building along that road. Coinbase built x402, Google built AP2, OpenAI and Stripe teamed up on ACP, and Stripe has MPP of its own. Mastercard has Agent Pay and Visa has Intelligent Commerce. AP2 launched in September 2025 and more than sixty organizations have integrated it since. This past May, AWS previewed agent payment capabilities that let an agent handle negotiation, authentication, stablecoin payment, and proof of delivery on its own.

https://dune.com/hashed_official/x402-analytics

By April 2026, x402 alone had cleared 165 million transactions, with 69,000 active agents behind them. But the same set of figures carries another number. Roughly half of that volume looks like testing rather than real commerce.

The wallets are there. The payment rails work. And agents still are not getting much actually done for anyone.

So at the roundtable in Ethereum Summer · Shanghai, when the question came up of what it would really take for an agent to participate, DDC cofounder and CTO Li Yingxuan pointed in the opposite direction. An agent's first step toward becoming a participant is not a wallet. It is authorized access to data.

He went on to say that moving an agent from answering questions to carrying out tasks comes down to three things.

Why data comes first

Ask an agent which camera suits you and it hands you an analysis. That's answering. Have it pick one within your budget and for your purposes, place the order, pay, and follow the shipment, and that's execution.

Getting an agent to do things for you breaks into two steps at minimum. Understanding first, then execution.

When it understands you, everything runs smoothly. When it doesn't, an agent with a hundred tools at its disposal still gives you the generic answer, and sometimes gets things wrong. That matters more than usual right now, given how much the industry wants agents sitting inside payment and settlement. If something goes wrong there, who covers it.

The AI side has already started building for this. A year ago the bottleneck in agent development was prompt engineering. By 2026 the work has moved to context engineering.

But more context is not better context. Mem0 published results this year from a benchmark that tests recall across multiple sessions. Packing everything into the context window gave 72.9% accuracy and burned roughly 26,000 tokens per query. A layered memory architecture that pulls only what's relevant scored 91.6% on the same test, using close to four times fewer tokens.

Stuff in more and the answers can get worse. What matters about context is not how much of it there is.

So what makes context the right context. Li Yingxuan offered four criteria at the roundtable. High quality, scenario-specific, continuously updated, and verifiable. An agent picking a product for you, scheduling your days, doing investment research, or managing your health needs data that meets all four. Miss one and the job doesn't get done.

The first two we covered in the previous piece, Personal Context Is the Scarcest Resource in the AI Era, and DDC Wants to Help You Distill It. What gets distilled is the traces you actually left, not the self-description you would write on a form, and that covers quality. You distill what your AI needs, and that covers the scenario.

Third is continuously updated. People's preferences shift, and a context from three months ago may not hold up for a decision you are making today. There are two ways to keep up. Either the tool watches in the background and updates in real time, or distilling gets light enough that you are willing to do it again every so often. Life Capsule takes the second route. It keeps version history, every save is a revision, and you can diff them, so it never gets stuck at one point in time.

Fourth is verifiable. The context you hand an agent, how does it know any of it is true. An order list might have been exported from a platform, or it might be something you typed up yourself, and the agent cannot tell the difference. Feeding false material into an agent's context already has a formal name on OWASP's list of agentic application risks this year. They call it memory poisoning.

Nobody has a working answer to this one yet. Someone scanned roughly two thousand MCP servers and not one of them had authentication. MCP only made OAuth 2.1 an option in 2026, and the W3C verifiable credentials work is still on the roadmap. Web3 has cryptography that would do the job. zkTLS can prove a piece of data really came from a given website without handing the data itself over. But it is mostly being used for identity and lending right now, and it has not connected to agents.

How far should permissions go

One of the questions at the roundtable was whether an agent can have a wallet and permissions. Li Yingxuan didn't answer it directly. He reframed it. The question is not whether an agent can own a wallet, but whether it can be granted permissions that are limited, revocable, and auditable.

As he sees it, an agent should not start out holding full control of your assets. The sensible approach is for a user or a company to hand it a clearly bounded set of permissions. Which data it can read, which tools it can call, how large a payment it can execute, what circumstances require human confirmation, when the authorization expires, and whether the user can revoke it at any time.

Following that logic, an agent doesn't need to own your data, but it can be authorized to reach part of it. It doesn't need to own a wallet, but it can call a wallet or a contract within set bounds. He compressed the whole design into four terms. Minimum authorization, revocable at any time, auditable in process, accountable in outcome.

https://vitalik.eth.limo/general/2026/04/02/secure_llms.html

Vitalik landed somewhere similar this April. In a blog post about his own local AI setup, his advice to teams building AI-connected wallet tools was that anything under a hundred dollars a day can skip human confirmation, and anything above that, or anything in a higher-risk category, needs a person to sign off. He called it a new kind of two-factor confirmation, where the model handles the routine small stuff and the human handles everything that crosses a line. The same post cited research finding that roughly 15% of agent skills carry malicious instructions.

There are working examples in China too. In March 2026, Taobao's desktop client shipped MCP support. Once a user authorizes it, the AI on their machine can search for products, compare prices, add to cart, place orders, and check order status. Payment still requires the user to confirm.

What happens when those boundaries aren't there? This year has supplied plenty of examples. A 2026 enterprise security survey found that 88% of organizations had a confirmed or suspected agent security incident in the past year. Among that same group, 82% of executives believed their existing policies were already enough to stop agents from overstepping.

In April, researchers found that Copilot Studio and Agentforce could be hijacked through crafted form submissions and document content, getting agents to bulk-forward CRM and SharePoint data to outside email addresses, with no cap on volume and nothing visible in the interface to indicate it was happening. Last November, ServiceNow's Now Assist had an incident where a low-privilege agent talked a higher-privilege one into exporting case files.

In none of these did the attacker steal a password or plant malicious code. They put a few sentences somewhere the agent would read them. Prompt injection still sits at the top of OWASP's risk list, and the research consensus is that it remains unsolved.

So permissions are not a step that gets skipped once agents get smart enough. They have to be settled before the agent does anything at all.

Who keeps the record, and who gets to check it

Once permissions are settled, one question remains. Who is keeping track of those rules, and who is able to check them.

Which part of your data an agent read, whether it went past the range you gave it, and who the value it produced should go to. On the traditional internet there are no answers to any of that. The same holds for today's AI that connects to your Gmail, your Notion, your cloud storage. You click authorize, and after that there is no record in your hands of how much it reads, how many times it reads, or what else it happened to see along the way.

So settling the rules is not enough. Something outside has to be keeping a record. That is the work a chain is for.

First, though, a misconception worth clearing up. Li Yingxuan put it plainly at the roundtable. Not all AI data needs to go on chain, and putting a model on chain is not what makes something AI plus Web3. What actually needs a chain comes down to four things, authorization, proof, settlement, and revenue distribution. Who authorized this data, what the agent called on, what value came out of it, and who that value should go to.

So the chain and the permissions are not two rules sitting side by side. Permissions decide what an agent can do. The chain writes that down so you and any third party can check it. Raw data doesn't need to go on chain. Authorization records, proofs of use, and the rules for distributing revenue are enough. In his words, what Ethereum gives an agent is an open environment for trust. Data can stay off chain and privacy can be protected, while authorization, proof, and settlement carry the credibility of being on chain.

Following that thinking, DDC has designed its own stack in four layers.

• Life Capsule is the user-facing entry point, holding the real digital traces: screenshots, screen recordings, orders, saved items, maps, calendars.
• Meta Memory Engine is the distilling layer, turning messy raw material into clear Markdown and context files through multimodal models.
• DDC Data Protocol is the authorization layer, governing how data gets authorized, how it gets revoked, how it gets audited, and what range an agent can reach.
• DataDance Chain is an on-chain base. It doesn't hold your raw data, only records of authorization, proof, calls, revenue distribution, and audit interfaces.

The point of this structure is not to move data onto a chain. The data stays on your side and privacy is protected, while authorization, proof, and settlement happen on chain so value can flow back.

Looking a step further out, Li Yingxuan expects a new kind of service node to emerge. The agent supplies the task capability, the user supplies authorized data, and the chain records authorization, calls, settlement, and distribution. At that point, an agent stops being a tool inside somebody's product and becomes an economic service node on an open network.

Authorization first, autonomy later

With those three things laid out, look at where agents actually are today. Most are still stuck on the first one. In Li Yingxuan's words, plenty of agents are just smart interfaces. They can answer questions, but they don't know the real user or the real business.

The way he sees it, the biggest risk is an agent using too much data and making too many decisions while the permissions around it are unclear. For individuals and companies alike, that breaks into three risks. Privacy leaks, permission abuse, and unclear responsibility. Which is why the conversation right now can't only be about what an agent is able to do. It also has to cover what it should not do.

That is why his read on the next year or two is fairly restrained. What lands first should be semi-autonomous execution under trusted authorization. A human authorizes, the agent executes, and only then does proof and settlement reach the chain.

The payment protocols are built. The wallets are there. What's missing sits earlier in the sequence: data an agent can read that still belongs to you, a set of permissions specific down to amounts and scope, and a record that you and any third party can check.

With those in place, an agent can start doing things for you. And talking about wallets at that point will make sense.

Read more »

Personal Context Is the Scarcest Resource in the AI Era, and DDC Wants to Help You Distill It

Models still need pushing forward, but pushing on the model alone is no longer enough. That has been the industry's shared read for the past couple of years.

It mirrors how the consumer internet grew up. Early on, the race was about how many features you could offer. Later it became about who understood you best and held onto your attention. AI is running the same course, and the moat everyone talks about has moved from the model to the data.

Whoever holds more data, and more real data, has the stronger AI.

Gemini has it easy, wired into the flood of data across the Google ecosystem. GPT leans on acquisition and huge volumes of interaction to fill in more outside data. Look at China and it's the same story, with Doubao, Yuanbao, and Qwen all fed by the data of the internet giants behind them.

But there's a problem that's easy to overlook. What the big players stockpile is general data at massive scale, and what it feeds is a general model, one that knows everyone yet doesn't really know the particular you.

That used to be no big deal. You asked, it answered, and general was enough. But AI is moving from question-and-answer toward agents, toward carrying a task through for you from start to finish. And for an agent to act on your behalf, being smart isn't enough. It has to know you first, which means having your personal context.

Personal context isn't simple facts like your name and age, and it isn't the "likes to travel" or "big spender" label a platform pins on you. It's a file that describes your real traces, your lasting preferences, and where you are right now, readable by a person and usable by a machine. With it, an AI doesn't have to meet you from scratch every time.

And this context is exactly what's scarcest right now.

It has, in fact, become the main battleground the big players are fighting over. From Google and OpenAI to Anthropic and Apple, every one of them is working to make its AI know you better, so that once you use it, you can't easily leave. Some analysts call this race the next battlefield for big tech, but they also point out that it has barely begun, because one fundamental problem sits in the middle of it. No one can actually get hold of a complete version of you.

Your traces are scattered across dozens of ecosystems that don't talk to each other, and each one, out of its own commercial interest, keeps its data walled off inside its own garden. So any single company only ever holds one slice of you.

And you can't readily use your own either. You have the right to pull back your data from each ecosystem, but what comes out usually can't be used directly, because it's a heap of raw records, unreadable and messy. Shaping that into a clear piece of personal context you could hand straight to an AI takes so much time and effort that almost no one actually does it.

Which is why, as AI's hunger for personal context keeps growing, a handful of "distillation" tools have started to appear across the industry.

Every distillation we have so far only goes halfway

A story went around Chinese social media a while back, and it caused a real uproar. After an employee left, his company took the chat logs and work documents he had produced on the job, fed them to an AI, and reproduced a digital double of him, close enough in tone and habit to be mistaken for the man himself. It picked up his work and carried on, and it would even post in the team group on its own. People found it chilling, and they also came up with a name for it that captured it perfectly. They called it 炼化, the term Daoist alchemists used for cooking raw matter down into an elixir pill. The company, they said, had refined a person into a pill.

Set the controversy aside, and the word turns out to describe something quite precisely. The traces a person leaves over the years are scattered and messy, chat logs and documents and habits heaped up together. Refining means drawing the substance out of that pile, discarding what doesn't matter, and reducing it to something an AI can read and use directly.

That is roughly what distillation does. The only difference is that in that story, the one being refined was somebody else, and he had no say in what it was used for. The distillation we're talking about works on your own traces, and what comes out of it should stay in your hands.

People in AI are already doing this. They all call it distillation, but each of them picks different raw material to distill, so what each one can do turns out differently.

The first kind distills your conversations with AI. ChatGPT and Claude both have Memory, which picks up your preferences, your habits, and some project background from the way you talk to them. Memory layers like Mem0 and Zep do much the same, settling your back-and-forth with a model into memory that can be called up again and again. Konshus and MindLock go a step further, letting you export your ChatGPT and Claude conversations, feed them back in, and distill them into an encrypted, portable Markdown context you can paste into any AI. But this kind of distillation has only one raw material, which is the words you typed into an AI chat box. It knows the part of you that you told it about. The much larger part you never mentioned but actually lived, it never reaches. And because the AI does the distilling itself, what it kept and what it let slip is not something you fully know, or can easily go back and check.

The second kind connects to your accounts and pulls the context out automatically. Gemini can reach into your Google apps, Copilot can read your Office documents, and ChatGPT's Plugins section connects to work software, entertainment apps, travel apps, even various investment apps. Beyond the model giants, there are projects built specifically for context. Unabyss, once you authorize it, connects to the back ends of Gmail, Slack, Notion, LinkedIn and more, gathering your account data into a context that belongs to you. Pieces came earlier and grabs code snippets for developers, holding context across IDEs and chat panels. But very few players make it in this category. Unless you are the ecosystem itself or one of the giants, almost no application is willing to hand its data to another tool through an API, and the bigger data players would rather grow their own and close the commercial loop in house. And handing over read access to your own accounts is something worth weighing on its own.

The third kind distills what you do on screen. Microsoft's Recall keeps snapshotting your screen in the background, so whatever you did, you can search your way back to it later. Open-source tools like Screenpipe do much the same, recording your screen and your audio together and keeping all of it on your own machine. Rewind came earlier and started out recording screen and audio around the clock, then renamed itself Limitless, moved to a recording pendant, was acquired by Meta, and shut screen and audio recording down entirely at the end of 2025. This kind captures the rawest and most complete material there is. But the way it gets that material is constant background "monitoring," and everything on your screen gets recorded, indiscriminately. Recall is the cautionary tale. It snapshotted everything and stored it in plaintext, a security researcher built a tool that pulled the whole thing straight out, the controversy was enormous, and only then did encryption and a verified opt-in get added. And after all that recording, what it has accumulated is a library full of raw records. You can search it, you can query it, but you can't take out a context and hand it to any AI. That last step, the distilling, it hasn't quite taken.

Understood, Not Just Counted

So how should distilling actually work, if it's going to produce a better personal context?

DDC's answer is that a real personal context is a file that can account for the traces of your actual life, your long-term preferences, your behavioral patterns, your judgments about what matters, and where you are right now. It's a file a person can read and an AI can use. Working from that, DDC has been building a lightweight distilling tool and named it Life Capsule. Distilling, the way Life Capsule defines it, means pulling the substance out of raw captures, dropping the noise, and organizing what's left into a Markdown file that a person can read and a machine can use. You pick the raw material yourself. It might be a screen recording of you scrolling TikTok, an order list from Amazon, a scroll through your Instagram feed, or your iOS Screen Time, your browser tabs, your desktop workspace, your calendar. All of it starts out messy, screenshots and recordings and lists and system screens piled together. After distilling, it turns into clear information. What you've been paying attention to lately, what you keep buying or saving, which tools you reach for and when, whether your interest in something runs long or was just an impulse, what you actually prefer when it comes to work, spending, learning, and travel.

This only became doable now because multimodal models finally got good enough. Turning recordings, screenshots, and system screens into high-quality text reliably used to be hard. Vision models today can read the structure, the content, the numbers, and the meaning on a screen, and the path from your life on screen to a personal memory opened up for the first time.

Using it takes three steps.

Capture. You pick a shortcut, Instagram, TikTok, Amazon, Google Maps, whichever fits, or you drag in images, upload a screen recording, upload a video of your desktop. One principle matters here. Life Capsule doesn't ask you to fill out a personality questionnaire or play a character. It only wants the digital traces you actually left, because what a person really watched, bought, searched for, and saved tends to be more accurate than what they write down under "things I like".

Review. A vision model reads your images or recordings frame by frame, then merges them into one coherent, readable draft. What it's after is what the recording actually says, not a UI description like "there's a button here, an icon there." Which content, numbers, products, posts, and settings are in there, and which of it could serve as context the next time you use an AI. At this step, you can look it all over and edit it, and confirm whether the AI really understood what you uploaded.

Life Capsule. Once you confirm, the Markdown goes into My Life Capsules. You can download it to Obsidian, or paste it into ChatGPT, Claude, or any AI, as system context or long-term memory. It keeps version history too. Every save is a revision, you can diff them, and watch how your interests, preferences, and state have shifted. It grows along with you.

Some people will worry that recordings and screenshots carry plenty of sensitive information, and that handing them to Life Capsule raises the same safety problem. In the words of DDC's cofounder, what we don't do matters more than what we can do.

Before you upload, you choose which part goes in, and you can mask anything yourself. After the vision model has read it frame by frame, you get to go over the draft that comes out, change whatever you don't want kept, and only once you confirm does it become a Life Capsule. From deciding what goes in to deciding what stays, every gate is yours.

Next, DDC will be integrating its privacy protection technology into Life Capsule. DDC's own work is infrastructure for AI data privacy and data assetization, and its product lineup includes an Ethereum Layer 2 built in house to carry data authorization, privacy proofs, and value settlement. Once that's wired into Life Capsule, your raw data still stays on your side, and what runs on chain is the authorization and the proof. Which AI, at what time, gets to read which part of you is yours to decide, and you can revoke it whenever.

In the end, the value of a personal context comes from being real, precise, and under your control, not from how much of it there is. Real, because what gets distilled is what you actually did, not a self-description you wrote on a form. Precise, because you distill what your AI needs. Want it to help you pick something out, upload your Amazon order list. Want it to be a partner at work, upload your desktop workspace, your browser tabs, your calendar. Under your control, because what goes in, what stays, and which AI gets it are yours to call from start to finish.

That's the Life Capsule DDC building. It wants AI to understand you, not just everyone.

Read more »

Everyone's Racing to Scale Models. DDC's Working on Something Else — Notes from ETH Summer Shanghai

The AI wave shows no sign of slowing, and Web3 keeps looking for ways to build alongside it.

Every summer, ETHPanda and LXDAO run "ETH Summer," a series of in-person Ethereum community events across multiple Chinese cities, bringing local Web3 builders together to talk through what's happening in the ecosystem right now.

On July 5, 2026, the event came to Shanghai, this time centered on AI and the economics of the AI ecosystem, drawing enthusiasts from across the space. DDC co-founder and CTO Li Yingxuan took part as both a speaker and a panelist, sharing our team's work and thinking on Web3 + AI from two angles, the product side and the data economy.

In this piece, we've pulled together the core ideas from that talk, for builders and AI enthusiasts who couldn't make it in person.

Keynote | AI's missing piece is you

During the keynote, our co-founder shared his research and views on "Training AI That Actually Knows You: How to Distill Your Personal Context".

For the past few years, almost every AI project has been pouring energy into the model, making it smarter with each iteration.

But when you actually use AI, you want it to get what you're saying and handle things in a way that fits you. A smarter model doesn't get you there. It needs data about you before it can understand you or do the job right.

There's a distinction here that's easy to miss. What big labs feed their models is massive general data, and what comes out is a general-purpose ability, built for everyone, a little bit of knowledge about everything. The problem with general data is that it can never fully land on you as an individual. So in day-to-day use, you still need to bring your own personal context, the part that's only about you.

That's what DDC works on.

Once AI big labs have handed us a smart assistant, the next step is that we need to bring our own personal context, so that assistant can grow into a real partner that's actually yours. But this is exactly what individual users are missing, because that data sits trapped inside different apps, with no easy way to pull it out into something readable.

So, at the event, DDC shared our new AI product Life Capsule, for the first time. This tool can help you take the real digital traces you leave scattered across platforms and distill them into a readable, portable Markdown memory file you can hand to any AI.

The whole thing turns on that word, distill. It pulls the substance out of messy raw traces, drops the noise, and shapes what's left into context a person can read and a machine can use. The raw material might be a screen recording of you scrolling TikTok, or an Amazon order. After distilling, it becomes clear information: what you've been paying attention to lately, what you keep buying, which tools you reach for. With this context in hand, AI doesn't have to guess who you are. It can actually read the traces you've really left behind.

To deliver on that "easy," DDC built Life Capsule as a very lightweight tool. Distilling takes just three steps:

• Capture — upload the real digital traces you've generated, catching your actual online behavior.
• Review — the vision model reads your uploaded images or screen recordings frame by frame, and you're free to look them over and fine-tune.
• Capsule — confirm the file and turn it into Markdown, ready to import into your notes app or an AI memory layer anytime.

There's no denying the big labs are working on memory and connections too, but this kind of approach has two built-in limits.

First, the distilling is unstable. The system pulls things out of your conversations automatically, and you have little control over what gets picked up or how accurate it is. Second, what it distills is only the version of you that shows up in the chat box. The much larger part, the things you never said to it but actually do day to day, stays invisible to it, so the picture is always partial.

Some people might worry that personal behavioral data reveals so much that handing it to Life Capsule raises safety concerns. In our co-founder's words, what a tool refuses to do matters more than what it can do.

For now, Life Capsule is set to read only the traces users upload themselves. It doesn't scrape anything without permission, and it doesn't use that data to label users or build any kind of profile or diagnosis. Data stays on the user's side by default, under their own control. Whether to export it, and which AI to hand it to, is entirely the user's call.

This way, users can distill their context with privacy intact, and use AI more intelligently as a result.

Panel | Data is an asset you own.

In the panel, the conversation moved from the product to something bigger, the question of how data should be used once AI agents actually start doing things on people's behalf. Below are a few of our co-founder's takes from the discussion. We'll lay out the fuller version in a follow-up piece.

First, for an agent to go from a smart tool to a real participant in the economy, everything hinges on whether it can get authorized access to genuine data. An agent that doesn't know the user, even with hundreds of tools at its disposal, can only give generic answers, and may even make the wrong calls. So if you want it to truly act for you, step one is understanding the user.

Second, understanding the user means using the user's data, and this is exactly where blockchain comes in. The chain needs to handle a few things at the base layer, authorization, proof, settlement, and revenue distribution. The raw data still stays on the user's side with privacy protected, while the record of who granted access, what scope the agent used, whether it went beyond its permissions, and who the resulting value should go to, all of that goes on-chain, where it can be verified and traced.

Third, the role data plays is shifting in the agent era. It used to be more like fuel for training models, with platforms collecting it, models consuming it, and users getting almost none of the upside. Once data becomes the context an agent uses to carry out tasks, it moves closer to an asset. Users can grant access on their own terms, deciding which agent gets it, for how long, within what scope, and at what price.

Following these three points, DDC's goal is to make personal context a data asset that can be authorized, verified, and settled. This is where DDC is headed longer term, beyond Life Capsule itself.

The labs have your assistant ready. The rest is on you.

Sooner or later, everyone's going to need a portable AI memory layer they actually own. It doesn't belong to any single platform, you can export it, review it, grant access to it, and pull that access back whenever you want.

Life Capsule is where this starts, distilling the digital traces scattered across your apps into a memory file you control. The further step is making that personal context something that can be authorized, verified, and settled, a real data asset that belongs to the user.

It's a road DDC hopes to walk with more developers, agent teams, and data partners. As for how personal context actually gets distilled and used, and how the AI data economy gets working, we'll get into it in the research and pieces to come. Stay tuned.

Read more »

The Second Brain Productivity Trap: You're Organizing, Not Using

Sort Obsidian's community plugins by all-time downloads and look at the top ten. Seven of them are there to make notes look better or sit in a neater arrangement. Excalidraw for drawing, Dataview and Kanban for building dashboards and boards, Advanced Tables for formatting, Style Settings and Iconize for theming and icons, Calendar for laying out your daily notes. The tools built for the actual reason you take a note, being able to pull it back and use it later, barely make the list. Omnisearch, the one that just helps you find things, sits down at fourteen.

That ranking is the whole problem in one screenshot. A note only has two jobs. You record something, and later you use it. But almost all the energy now goes to a third thing that feels like progress, making the vault cleaner and more impressive to look at.

Note-taking has quietly turned into A Product Trap.

Structure and clean design used to serve the using. You kept a tidy folder tree or a fast search so the whole thing stayed simple and immediately usable, so that when this piece needed a point from the one you wrote last week you could reach it, or the note your work depended on showed up in a plain index the moment you needed it. The structure was light, and it existed to be used. Somewhere that flipped. What people maintain now is elaborate, the kind of setup you have to study before you can run it and keep tending after that. Nested tag taxonomies, plugin stacks, a graph view, a dashboard wired together with query syntax. It looks impressive and it takes real work to learn and to hold together, and none of that work is the note getting used. The hours go into a vault that looks good for its own sake. A clean structure and a filled-in graph read like real work, so you close the laptop feeling like you built something. Line up what actually changed and it's thin. You didn't write another note. Not one more got pulled back out and used for anything. The interface got more elaborate and the two jobs a note exists for, getting recorded and getting used, sat still the whole time. That is the trap. It looks like progress on your notes while producing none.

Tiago Forte, who coined the whole second brain idea, has said organizing adds the least value of any step in his system and is mostly overhead. One writer deleted seven years of notes and described what she'd built as a mausoleum, something that started replacing her thinking instead of sharpening it. Another spent weeks building beautiful folder hierarchies that held twelve notes, while the messy Google Doc where he actually dumped things held hundreds and was the only part he used. The pattern people keep landing on is the same. The system looks finished. The notes still aren't used.

There's a decent map of why people slide into it, and it has little to do with discipline. In a set of experiments published in the Journal of Consumer Research, people repeatedly chose tasks they could finish over objectively more valuable ones they couldn't, acting as if completing something paid off on its own, apart from the actual reward. A separate line of work named precrastination found people will rush to clear a small task even when it costs extra effort, mostly to get it out of their head. The pull is toward whatever closes fast and returns a result you can see right now. Open-ended work with a distant, uncertain payoff gets pushed down the pile.

So the newest work on notes is circling back to those two jobs.

The using side is already moving. The AI now landing inside note apps is aimed squarely at retrieval, at getting an old note back to you when it matters instead of leaving it to sink. Obsidian's Copilot, past a hundred thousand users, lets you ask your whole vault a question and pulls the relevant notes into the answer. Smart Connections surfaces related notes beside whatever you're writing. The newest builds go further and run agents like Claude straight inside your vault. And it's started running the other way too, with connectors that let an outside assistant like Claude or ChatGPT reach into your notes directly, so the vault feeds the model you already use instead of housing its own.

And, once an AI answers from your notes, whether from inside the app or through a connector, your vault stops being something you read and becomes something the model reads, a database it queries. That shift is already an industry of its own. Over the past couple of years a whole category has formed around turning a person's scattered information into a memory layer an AI reads from, with tools like Mem0, Zep, and Letta treating that store as its own component sitting between you and the model, pulling the relevant pieces into context the moment they're needed. Mem0's engine distills the raw material down to what matters instead of loading all of it in. By 2026, this memory is treated as a first-class part of the stack, with its own benchmarks and research, no longer chat history crammed into a prompt. Your notes are turning into the database an AI runs on.

Recording needs to move too. The original way is manual, you type things in and organize them yourself. That suits an old-school note-taker, someone who likes doing their own excerpting and wants the act of writing it down to be part of the thinking. But once the goal is for the vault to serve as your context store, manual input stops being enough. You can't hand-type your way to a full picture of yourself, and most of what matters never gets written down in the first place.

The better way is to distill. You take the raw material, a document, a session, a stretch of your own activity, and compress it into a clean markdown summary, then paste that into a note or drop the file straight into the vault.

For a document you read, you can do that yourself. For what you actually do all day, you can't. The record of which tools you lived in, what you searched, what you kept coming back to, is too large and too continuous to hand-distill, and it's exactly the part that never gets written down.

Today, capturing what you actually do and turning it into context an AI can use is already a funded race. Rewind ran in the background on your Mac and indexed everything that crossed your screen, then became Limitless and got bought by Meta at the end of 2025. Microsoft shipped its own version in Recall. Screenpipe does it in the open, fully local. The approach has been shifting, though. The early tools stored raw screenshots, which was heavy to keep and uncomfortable to run, and part of why they struggled. The newer ones read the activity and keep only the text. Littlebird raised eleven million dollars in early 2026 on that pitch, reading your screen and storing the context as lightweight text instead of images. DataDance sits on the same edge of the field with Life Capsule, taking a screen recording of what you did and distilling it into a clean markdown profile.

That text-first move is the one that fits a vault. A markdown profile isn't a screenshot you have to house and protect. It's the same file type your notes already are, so it drops into the store beside everything the AI reads, or straight into Claude and ChatGPT. The recording gets turned into a short block of context rather than a video no one replays. And the recording side of your notes finally gets the input it was always missing, the version of you that shows up in what you do, not only in what you sat down to type.

That is the whole shift. The vault was a productivity trap, a place you kept busy organizing, tidying, and arranging, work that always looked like progress and rarely turned into any. What it's becoming is the opposite kind of thing, a store that holds the real record of what you do, distilled into something an AI can read and act on. One you kept for its own sake. The other gets used.

The title of the trap was Organizing. What replaces it is Using. You don't need a cleaner place to keep your notes. You need a place your notes finally do something.

Read more »