<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by ben thomasson on Medium]]></title>
        <description><![CDATA[Stories by ben thomasson on Medium]]></description>
        <link>https://medium.com/@ben.thomasson?source=rss-1043371988b7------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*i0YcQm1rMA39CNYROtn_Pg.jpeg</url>
            <title>Stories by ben thomasson on Medium</title>
            <link>https://medium.com/@ben.thomasson?source=rss-1043371988b7------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 14:20:47 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ben.thomasson/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Keep a Diary]]></title>
            <link>https://medium.com/@ben.thomasson/keep-a-diary-c36cf36b805c?source=rss-1043371988b7------2</link>
            <guid isPermaLink="false">https://medium.com/p/c36cf36b805c</guid>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[ben thomasson]]></dc:creator>
            <pubDate>Mon, 29 Jun 2026 05:01:49 GMT</pubDate>
            <atom:updated>2026-06-29T13:33:34.837Z</atom:updated>
            <content:encoded><![CDATA[<p>You should keep a diary.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3y_5BDSOW4Gk0He2.png" /></figure><p>Not the kind you write before bed about your feelings. A research diary. A thinking diary. A diary of what you learned, what you decided, and why.</p><p>If you’re working with LLMs — and increasingly, who isn’t — this is not optional advice. It’s load-bearing infrastructure.</p><h3>The Problem You Already Have</h3><p>You’re in the middle of a project. You had a conversation with an LLM last Tuesday where you worked through a tricky design decision. You considered three approaches, rejected two for specific reasons, and chose the third. Today you can’t remember which approach you chose or why you rejected the others. The LLM certainly doesn’t remember — that context is gone.</p><p>So you have the same conversation again. You re-derive the same conclusions. You might even make a different decision this time, not because you have new information but because you forgot the old information. The LLM helps you think it through again, confident and fluent, as if for the first time. Because for it, it is.</p><p>This is the sawtooth pattern. You build up reasoning, lose it, rebuild it, lose it again. Each cycle wastes the work of the previous one. Doyle identified this problem in 1979: “discarding derivation records may be condemning ourselves to continually rederiving information.” He was talking about AI systems, but he was also talking about you.</p><h3>The Solution Is Older Than Computers</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*W2eHb5ao0InGp5r6.png" /></figure><p>Scientists have always kept lab notebooks. The practice isn’t about memory — scientists have good memories. It’s about <em>justification</em>. When you write down what you observed, what you concluded, and why, the reasoning becomes inspectable. You can return to it months later and evaluate whether the conclusion still holds given what you know now. You can share it with a collaborator who wasn’t in the room. You can catch your own mistakes.</p><p>The lab notebook is external memory — knowledge that lives outside your head, carries its justifications, and lets you trace how you got from observation to conclusion. Scientists didn’t need LLMs to realize this was essential. They needed it because human memory is lossy, human reasoning is fallible, and important work requires a record that survives both.</p><h3>What Changes with LLMs</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-6yMfr_EKc6WU4bn.png" /></figure><p>Two things change.</p><p>First, the throughput of thinking goes up dramatically. An afternoon with an LLM can cover ground that would have taken weeks of solo work. You explore more options, analyze more data, consider more angles. The volume of reasoning that needs to be captured per unit of time is much higher than in a traditional lab notebook.</p><p>Second, the LLM is a second thinker with zero memory. Every session starts from scratch unless you provide context. The diary is how you provide that context. Without it, you’re re-briefing a brilliant but amnesic collaborator every morning. With it, you hand them the notebook and say “read this, then let’s continue.”</p><h3>What Goes in the Diary</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Kxt7AEwlzglWALMo.png" /></figure><p>An entry is a unit of thinking. Not a transcript of a conversation — that’s too raw. Not a polished paper — that’s too refined. An entry captures:</p><ul><li><strong>What you learned.</strong> An experiment result. A finding from reading code. Something a reviewer pointed out.</li><li><strong>What you decided.</strong> Which approach you chose and why. Which alternatives you rejected and why.</li><li><strong>What surprised you.</strong> The unexpected finding. The assumption that turned out to be wrong. The connection you didn’t see coming.</li><li><strong>What you don’t know yet.</strong> The open question. The experiment you need to run. The thing that would change your mind.</li></ul><p>Each entry is dated. Each entry stands alone — you can read it cold and understand what it says without reading every entry before it. Each entry is specific enough that it’s useful six months later when you’ve forgotten the context.</p><h3>The Practice</h3><p>I’ve written 296 entries over four months using a simple CLI tool ( entry create &quot;title&quot;). Each entry is a markdown file in a date-organized directory. The LLM can read them, search them, reference them. When I start a new session, the LLM can read the last few entries and pick up where we left off.</p><p>Some entries are small — a paragraph capturing a single insight. Some are large — a full analysis of an experiment with data tables and implications. The important thing is not the size but the habit. If you learned something today that you’d need to re-derive tomorrow, write it down.</p><p>The entries compound. A single entry is a note. A hundred entries are a knowledge base. When I needed to write a paper about confabulation in LLMs, the entries from two months of experiments were the raw material — the observations, the results, the failed hypotheses, the corrections. The paper was assembled from entries, not invented from scratch.</p><h3>The Diary as Infrastructure</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0pfh6I2pYtgW0KkA.png" /></figure><p>Here’s what I didn’t expect: the diary changes how you think, not just what you remember.</p><p>When you know you’re going to write an entry, you think more carefully. You distinguish between “I feel like this is true” and “here’s the evidence.” You notice when you’re making an assumption. You catch yourself hand-waving. The act of writing for a future reader — even if that reader is yourself in two weeks — forces clarity.</p><p>And when you compose entries into larger artifacts — papers, blog posts, presentations — you’re not starting from a blank page. You’re selecting and arranging material that’s already been thought through. The heavy lifting happened in the entries. The composition is assembly, not invention.</p><p>You don’t sit down to write a paper. You sit down with four months of entries and realize the paper is already there, distributed across dozens of dated observations. You just have to find the shape.</p><h3>Start Today</h3><p>You don’t need special tools. A directory of markdown files organized by date works. I use a CLI ( entry create &quot;title&quot;) that creates a file at entries/YYYY/MM/DD/title-slug.md with a date header. Anything that creates a dated text file will do.</p><p>Write one entry today about whatever you’re working on. What did you learn? What did you decide? What surprised you? What don’t you know yet?</p><p>Do it again tomorrow. And the day after.</p><p>In a month you’ll have a knowledge base. In three months you’ll have the raw material for whatever you need to write. In six months you’ll read an entry from today and be grateful you wrote it, because you will have forgotten everything it contains.</p><p>Keep a diary.</p><p><em>Originally published at </em><a href="https://benthomasson.com/ai/research/workflow/keep-a-diary/"><em>http://benthomasson.com</em></a><em> on June 29, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c36cf36b805c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LLMs Have No Memory of Time]]></title>
            <link>https://medium.com/@ben.thomasson/llms-have-no-memory-of-time-70a63d43d3dc?source=rss-1043371988b7------2</link>
            <guid isPermaLink="false">https://medium.com/p/70a63d43d3dc</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[claude]]></category>
            <dc:creator><![CDATA[ben thomasson]]></dc:creator>
            <pubDate>Sun, 22 Feb 2026 05:00:00 GMT</pubDate>
            <atom:updated>2026-02-22T13:47:18.321Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3m598uK5vneKYoRGTA6BiQ.png" /></figure><p>Ask Claude what day it is and it’ll tell you. Ask it whether the thing it read five minutes ago is newer than the thing it read an hour ago, and it has no idea.</p><p>This isn’t a bug. It’s a fundamental property of how LLMs work. Everything in the context window is equally “now.” There’s no before and after. No newer and older. A claim from January 2025 and a claim from February 2026 sit side by side in the context, and the model treats them as equally current.</p><p>I’ve been running a multi-agent AI research programme — six AI agents across seven repositories — and this one property caused more problems than anything else.</p><h3>What Goes Wrong</h3><p>When you give an AI agent a role definition file (I use CLAUDE.md), that file becomes ground truth. The agent reads it, believes it, and acts on it. The problem: the file doesn&#39;t know when it was last updated. Neither does the agent.</p><p>I ran an audit of all six role definition files in my research programme. Every single one was stale:</p><ul><li>The researchers repo’s README said “the framework hasn’t derived the formula” — but an entry from three days earlier showed it had</li><li>The reviewer’s role definition didn’t mention breakthroughs from the researcher’s repo made the same week</li><li>The PI’s own file contradicted its own entries — the tensor sector was marked “resolved” in the role definition but classified as “falsified” in an entry from the same repo</li></ul><p>Six files. Six agents. All operating on outdated beliefs. None of them knew it.</p><p>This is what happens when you build a system with no sense of time.</p><h3>The Classical AI Version of This Problem</h3><p>It turns out this isn’t a new problem. In 1969, McCarthy and Hayes identified what they called the <strong>frame problem</strong>: how does a reasoning system know what changes when something happens, and what stays the same?</p><p>The practical compromise is called the <strong>sleeping dog strategy</strong> — assume everything stays the same unless you explicitly change it. It works most of the time. When it fails, there’s no detection mechanism. The dog sleeps through changes it should have noticed.</p><p>That’s exactly what my agents were doing. They assumed their role definitions were current. Nobody woke the sleeping dogs.</p><p>The belief revision literature (Alchourrón, Gärdenfors, and Makinson, 1985) makes a useful distinction: <strong>revision</strong> is updating beliefs because you got new information about a static world. <strong>Update</strong> is changing beliefs because the world itself changed. A role definition going stale because the research moved forward is an update problem, not a revision problem. They require different operations. My agents had neither.</p><h3>The Fix: Encode Time in the Filesystem</h3><p>The solution I landed on is embarrassingly simple: put the date in the directory path.</p><p>Every entry lives at entries/YYYY/MM/DD/filename.md. The path <em>is</em> the timestamp. An agent reading the filesystem can see that entries/2026/02/22/ is newer than entries/2026/02/21/ without parsing any content. Time ordering is structural, not semantic.</p><p>This works because agents read paths structurally but read content atemporally. Put a date inside the file and the agent treats it as text. Put the date in the path and it’s part of the navigation structure.</p><p>I built a small CLI tool called to enforce this:</p><p>Three things happen automatically:</p><ol><li>The date directory gets created from today’s date</li><li>The title gets slugified (no spaces — agents love creating filenames with spaces, which breaks everything downstream)</li><li>A standardized template gets filled in (Date, Time, Overview, Details, Next Steps, Related)</li></ol><h3>Why Not Just Use Timestamps in Files?</h3><p>I tried that first. It doesn’t work.</p><p>When an AI agent reads a file that says **Date:** 2026-01-15 at the top, it processes that as text content, not as temporal metadata. The model has no mechanism to compare that date against another file&#39;s date and conclude &quot;this one is newer.&quot; It can do the comparison if you explicitly ask, but it doesn&#39;t do it automatically when deciding what to believe.</p><p>Directory paths are different. When an agent lists files or navigates a filesystem, the path structure is part of how it locates information. An agent looking at entries/2026/02/22/ and entries/2026/01/15/ processes the path hierarchy as navigation - and navigation has inherent ordering.</p><p>The directory structure also makes git log, ls, and find work naturally. You don&#39;t need special tooling to answer &quot;what happened last week?&quot; - it&#39;s just ls entries/2026/02/15/ through entries/2026/02/22/.</p><h3>The Right Time Unit Depends on the Domain</h3><p>Calendar days work for research journals, project logs, and meeting notes. But I also tested this on an automated software development pipeline that runs in minutes, not days. All the entries landed in a single day directory — useless.</p><p>The fix: use the right temporal unit for the domain.</p><ul><li><strong>Research programmes:</strong> calendar days (entries/YYYY/MM/DD/)</li><li><strong>SDLC pipelines:</strong> iteration numbers (entries/iteration-1/, entries/iteration-2/)</li><li><strong>CI systems:</strong> build numbers (entries/build-42/, entries/build-43/)</li></ul><p>The principle — impose temporal structure externally because the model can’t maintain it internally — is constant. The granularity varies.</p><h3>An Unexpected Finding: Entries as Specs</h3><p>Something I didn’t design for: entries turned out to work as cross-session coordination artifacts.</p><p>In one case, I wrote a to-do list as an entry with six prioritized items for improving an automated SDLC pipeline. I committed it to git. A separate Claude session — with zero shared context, no conversation history, no handoff — read the entry and implemented all six items in 67 minutes. 371 lines of code. Plus a follow-up bugfix for an edge case the spec didn’t mention.</p><p>The entry worked as a spec because it had:</p><ul><li>Specific file and line references (the agent didn’t need to explore)</li><li>Concrete examples (verdict block format, exit gate logic, prompt text)</li><li>Priority ordering matching natural implementation order</li><li>Rationale explaining <em>why</em>, not just <em>what</em></li></ul><p>Two sessions coordinated on a 371-line change through a single markdown file in a date-organized directory. No shared context, no handoff meeting. The filesystem was the coordination mechanism.</p><h3>What I Learned</h3><p>After hundreds of commits across 7 repositories:</p><ol><li><strong>Models don’t have time. You have to give it to them.</strong> Timestamps in file content are decoration. Timestamps in directory structure are functional.</li><li><strong>Stale beliefs are the silent killer.</strong> Nobody notices when a role definition goes stale. The agent keeps working confidently on outdated information. An audit found 100% of role definitions had staleness issues.</li><li><strong>The right time unit matters.</strong> Calendar days for humans. Iterations for pipelines. Build numbers for CI. Match the granularity to the workflow.</li><li><strong>Entries are more than documentation.</strong> They’re coordination artifacts, specs, audit trails, and temporal anchors. Once you have time-ordered records that agents can navigate structurally, they find uses you didn’t plan for.</li></ol><h3>Try It</h3><p>Or just create the entries/YYYY/MM/DD/ directories yourself. The tool enforces conventions but the principle works with any filesystem.</p><p>The code is at <a href="https://github.com/benthomasson/entry">github.com/benthomasson/entry</a>. Zero dependencies. Three commands.</p><p><em>This is the first in a series on belief management for AI agents, based on a year-long multi-agent research programme. Next up: what happens when your agents start disagreeing with each other — and neither of them knows it.</em></p><p><em>Originally published at </em><a href="https://benthomasson.com/ai/tools/knowledge-management/llms-have-no-memory-of-time/"><em>http://benthomasson.com</em></a><em> on February 22, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=70a63d43d3dc" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>