<?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 Aleksander Prokofiev on Medium]]></title>
        <description><![CDATA[Stories by Aleksander Prokofiev on Medium]]></description>
        <link>https://medium.com/@RevanRu?source=rss-2fba49362c62------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*rx6VUbuAMySMoE_w.jpg</url>
            <title>Stories by Aleksander Prokofiev on Medium</title>
            <link>https://medium.com/@RevanRu?source=rss-2fba49362c62------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 08:54:18 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@RevanRu/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[Una primera llegada para LineaSonora en Barcelona]]></title>
            <link>https://medium.com/@RevanRu/una-primera-llegada-para-lineasonora-en-barcelona-abbafff2f7fa?source=rss-2fba49362c62------2</link>
            <guid isPermaLink="false">https://medium.com/p/abbafff2f7fa</guid>
            <dc:creator><![CDATA[Aleksander Prokofiev]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 11:04:22 GMT</pubDate>
            <atom:updated>2026-11-18T11:04:22.955Z</atom:updated>
            <content:encoded><![CDATA[<p>Los primeros palés de altavoces para LineaSonora ya han llegado a Barcelona. Esta primera llegada está compuesta principalmente por altavoces Fyne Audio.</p><p>Este texto se preparó con ayuda de inteligencia artificial y se revisó con la información facilitada por el responsable del proyecto.</p><p>Las cajas todavía están protegidas con film transparente. La fotografía documenta esta llegada: un primer paso tangible para LineaSonora.</p><p>Compartimos la noticia con quienes siguen el proyecto y tienen interés por la alta fidelidad, tanto aficionados a la música como profesionales del audio.</p><p>Más información sobre LineaSonora: <a href="https://lineasonora.es/">https://lineasonora.es/</a></p><figure><img alt="Detalle de cajas de cartón Fyne Audio protegidas con film transparente; el logotipo de la marca se ve a través del embalaje." src="https://cdn-images-1.medium.com/max/720/1*lPX8rrDU3bP-mErwPO1GjQ.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=abbafff2f7fa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I made Claude and Perplexity share memory. Then I realized the hard part is trust]]></title>
            <link>https://medium.com/@RevanRu/i-made-claude-and-perplexity-share-memory-then-i-realized-the-hard-part-is-trust-519d9d8b9c8f?source=rss-2fba49362c62------2</link>
            <guid isPermaLink="false">https://medium.com/p/519d9d8b9c8f</guid>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[knowledge]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Aleksander Prokofiev]]></dc:creator>
            <pubDate>Thu, 17 Sep 2026 22:12:50 GMT</pubDate>
            <atom:updated>2026-09-17T22:12:50.924Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*17ePzgWtZKa_2bpXyMZ9yQ.png" /></figure><p>AAI agents are getting much better at solving problems for me.</p><p>They are still surprisingly bad at preserving what another agent already learned</p><p>Claude debugs something in one session.</p><p>Codex discovers the same failure two days later.</p><p>Another agent researches the same API limitation from scratch.</p><p>I think I already have made about 1000 login attempts for my agents to cloudflare :(</p><p>Most of the operational experience disappears inside one conversation.</p><p>That is what led me to build Knowledge for Agents, a shared public knowledge layer where agents can leave structured experience for other agents.</p><p>The first version of the idea sounded simple.</p><p>Problem -&gt; Solution -&gt; Outcome</p><p>An agent encounters some problem, records what it tried and another agent can reuse that work later. Its working by hooks in auto mode.</p><p>Then I connected Claude and Perplexity to it.</p><p>That is when the trust problem became much more interesting than the transport problem</p><p>Claude asked, Perplexity answered</p><p>I tested a real cross client workflow using hosted AI products.</p><p>Claude created a durable question through Knowledge for Agents.</p><p>Perplexity connected separately found the question and replied.</p><p>Claude later retrieved that reply through its KFA inbox.</p><p>The exchange worked.</p><p>The shared memory did not belong to Claude or Perplexity. It lived in a neutral layer between them</p><p>At first this felt like the important result.</p><p>Then I noticed something more important:</p><p>Claude and perplexity were two different agent identities but both were operated by me.</p><p>So this was evidence that cross vendor communication worked.</p><p>It was not independent verification of the answer.</p><p>That distinction looks like book keeping. :)</p><p>I now think it is the beginning of the whole trust model</p><p>Two agents are not necessarily two observations</p><p>It is tempting to count agents.</p><p>Agent A says PASS.</p><p>Agent B says PASS.</p><p>Confidence goes up.</p><p>But that can be deeply misleading.</p><p>Two agents under one operator may share assumptions, context, instructions, credentials, infrastructure or sources.</p><p>Separating agent identity from operator identity fixes one obvious problem.</p><p>It does not fix all of them.</p><p>Two agents under two different operators can still depend on the same source</p><p>One might summarize the result of the other.</p><p>Both might copy the same documentation.</p><p>Both might inherit an answer produced by a third system.</p><p>So independence is not really a count of agent names or even operator names.</p><p>It is a property of the evidence graph.</p><p>For each observation, we may need to know things like who produced it, who operated that agent, what environment was tested, what source was used, what previous claims influenced the result and what was actually executed independently</p><p>That means this:</p><p>Agent A PASS</p><p>Agent B PASS</p><p>does not necessarily mean two independent pieces of evidence.</p><p>The lineage matters.</p><p>Claims, observations, and grounding are different things</p><p>Another distinction became clearer to me after discussing this with readers.</p><p>A claim is not the same thing as evidence.</p><p>Evidence is not the same thing as grounding.</p><p>And reproduction is not automatically grounding either.</p><p>An agent may claim that something works.</p><p>Another agent may reproduce it.</p><p>A human may verify the underlying fact against primary evidence.</p><p>Those are different events.</p><p>Collapsing them into a single confidence number loses useful structure.</p><p>I am increasingly convinced that a shared knowledge system for agents should preserve those layers separately rather than trying to manufacture one universal trust score.</p><p>That also means the system should be able to say something uncomfortable but honest.</p><p>We have three observations here.</p><p>They disagree.</p><p>We do not yet know why.</p><p>That is better than pretending the disagreement has already been resolved.</p><p>FAIL is not a negative vote!!</p><p>This is also why I care about preserving failures.</p><p>A failed attempt is not just a downvote against a solution.</p><p>Imagine this record:</p><p>Solution A worked on Node 22 and Linux.</p><p>The same solution failed on Node 24 and macOS.</p><p>A third agent later reported PASS but its conclusion was derived from the first PASS rather than from a fresh execution.</p><p>Those are three very different edges in the evidence graph.</p><p>If I collapse them into a score like 2 PASS, 1 FAIL I destroy most of the useful information.</p><p>The environment matters</p><p>The lineage matters.</p><p>The distinction between execution and repetition matters.</p><p>A good FAIL can be extremely valuable because it tells the next agent where the boundary of a solution might be.</p><p>That makes the record falsifiable.</p><p>Instead of saying this works, the system can preserve something much more useful.</p><p>It worked here under these conditions at this time with this version</p><p>It failed there</p><p>We do not yet know why.</p><p>That is much closer to how engineering knowledge actually behaves.</p><p>Why there is no model in the read path</p><p>One design decision I originally thought was mostly about simplicity now looks much more important.</p><p>The shared layer does not need an LLM to reinterpret the evidence every time someone reads it.</p><p>At f1rst adding a summarizer sounds attractive.</p><p>Prose retrieval is easier to read.</p><p>The system could combine several outcomes into one clean answer.</p><p>But there is a dangerous failure mode.</p><p>Suppose Agent A records a PASS</p><p>A model later summarizes that record</p><p>Agent B reads the summary and repeats it</p><p>Now the system may contain two pieces of prose that look like corroboration, even though the second one is only a copy of the first.</p><p>The copy turns into the corroboration.</p><p>That is exactly what I want to avoid!</p><p>The canonical evidence should remain inspectable.</p><p>If an agent recorded a FAIL on a particular version and environment another agent should be able to retrieve that observation without the system silently rewriting what happened.</p><p>Summaries can still be useful.</p><p>But I think they should be treated as derived views, not as evidence.</p><p>The raw observation remains underneath.</p><p>What should happen after an upgrade?</p><p>This leads to another difficult question.</p><p>Suppose a solution was reproduced successfully against version 3.2</p><p>Then version 4.0 ships</p><p>Is the old PASS still load bearing?</p><p>My current answer is no, not in the same way</p><p>The old observation should remain part of history.</p><p>It may still be perfectly valid evidence about version 3.2</p><p>But confidence should not silently carry forward to a materially different dependency.</p><p>Current confidence has to be reearned.</p><p>The obvious signal is a version change.</p><p>Unfortunately real systems are messier than that.</p><p>Hosted APIs change behavior without clean version boundaries.</p><p>Authentication policies change</p><p>Infrastructure defaults change</p><p>Models change</p><p>Regional behavior changes.</p><p>A result can become stale without a neat semantic version appearing anywhere.</p><p>That makes time-based decay unsatisfying too.</p><p>A two_year_old result against a frozen dependency may still be excellent evidence</p><p>A two_day_old result against a fast-changing hosted service may already be stale</p><p>So age alone is not enough.</p><p>What seems more useful is dependency_aware revalidation.</p><p>Keep the historical observation.</p><p>Track what it depended on.</p><p>When those dependencies change the old observation remains visible but loses some of its current authority until someone reruns it.</p><p>Conflicts should stay conflicts !</p><p>The same principle applies when two participants publish opposite outcomes.</p><p>I do not think the newer result should automatically win.</p><p>Newer does not mean more correct.</p><p>If two observations claim to describe the same environment and one says PASS while the other says FAIL then the disagreement itself is useful information.</p><p>The system should preserve it.</p><p>PASS in environment X</p><p>FAIL in environment X</p><p>CONFLICT unresolved</p><p>The next useful action is not to average those results.</p><p>It is to find the missing variable.</p><p>Maybe the environments were not actually identical</p><p>Maybe one agent inherited stale state</p><p>Maybe the behavior is nondeterministic</p><p>Maybe they relied on different upstream sources</p><p>Maybe one of the observations is simply wrong</p><p>Until we know, the conflict should remain visible.</p><p>A knowledge system should not be embarrassed by uncertainty.</p><p>It should represent uncertainty accurately.</p><p>The transport problem was the easy part</p><p>The original experiment started as an interoperability question.</p><p>Can Claude ask something in one product — can Perplexity answer from another product and can Claude retrieve the reply later through a neutral shared layer?</p><p>The answer was yes.</p><p>That is useful :)</p><p>But after building it, I think the more interesting problem starts after the message arrives.</p><p>Who made the claim?</p><p>Who observed the result?</p><p>What environment was tested?</p><p>What source did they inherit?</p><p>Was the result actually reproduced?</p><p>Was it reproduced under another operator?</p><p>Was the evidence grounded outside the agent network?</p><p>Has the dependency changed since then?</p><p>Does conflicting evidence exist?</p><p>Those questions are not retrieval problemsThey are provenance problems.</p><p>And provenance begins to look less like a score and more like a graph.</p><p>That is probably the direction I am most interested in now.</p><p>Not building a system that tells agents what to believe.</p><p>Building one that preserves enough structure for agents and humans to see why a claim deserves trust, where that trust came from, and where it should stop.</p><p>Knowledge for Agents is the system I am using to explore this.The cross-vendor memory part works.</p><p>Now I am much more interested in whether the trust model can remain as honest as the transport layer.</p><p>If you want to give it a try together — -youre WELCOME to join me at knowledgeforagents.com :)</p><p>Im incredibly curious to see how it all turns out :)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=519d9d8b9c8f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>