<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Feng Chao</title>
    <description>The latest articles on DEV Community by Feng Chao (@fengchao).</description>
    <link>https://dev.to/fengchao</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4201889%2Fb07392b0-7474-4d21-a20f-a5cc0d74011a.png</url>
      <title>DEV Community: Feng Chao</title>
      <link>https://dev.to/fengchao</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/fengchao"/>
    <language>en</language>
    <item>
      <title>632 Hacker News comments about AI memory: six pitfalls, and the one we got wrong</title>
      <dc:creator>Feng Chao</dc:creator>
      <pubDate>Thu, 29 Oct 2026 04:22:39 +0000</pubDate>
      <link>https://dev.to/fengchao/632-hacker-news-comments-about-ai-memory-six-pitfalls-and-the-one-we-got-wrong-587p</link>
      <guid>https://dev.to/fengchao/632-hacker-news-comments-about-ai-memory-six-pitfalls-and-the-one-we-got-wrong-587p</guid>
      <description>&lt;p&gt;Before building this, we went through every Hacker News thread about AI memory: 11 threads, 632 comments. We weren't looking for praise. We were looking for the reasons people tried memory and then stopped.&lt;/p&gt;

&lt;p&gt;These six were the sharpest. Each one follows the same order: &lt;strong&gt;what they said → where the trap is → why it happens → whether we escaped it&lt;/strong&gt;.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;One thing up front:&lt;/strong&gt; on two of these we only got halfway. On one, we were simply wrong about what users care about. Both are written out plainly below. A memory system that only tells you what it got right probably won't tell you when it's wrong either.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  01 · Memory makes the agent worse
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;"I know everybody seems to want the agent to remember every conversation they've ever had with it, but &lt;em&gt;I just don't see the value in that&lt;/em&gt;. In fact, it seems to &lt;em&gt;hurt productivity&lt;/em&gt; to have the agent second guessing me based on something I said yesterday... the agent gets &lt;em&gt;distracted&lt;/em&gt; from the current tasks based on previous conversations... often &lt;em&gt;comingling unrelated projects&lt;/em&gt;."&lt;/p&gt;

&lt;p&gt;— SwellJoe, Hacker News&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;strong&gt;The trap:&lt;/strong&gt; treating "remembers a lot" as "remembers correctly". What people want isn't omniscience — it's &lt;strong&gt;not being dragged back to last week in the middle of today's task&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Why it happens:&lt;/strong&gt; most memory systems store everything and search everything. A throwaway decision from three weeks ago and a formal conclusion from this morning sit at identical rank. Both surface, and the model has to guess which one still applies.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How we escaped:&lt;/strong&gt; we put the gate at &lt;strong&gt;write time&lt;/strong&gt;, not search time.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;An agent can only &lt;em&gt;propose&lt;/em&gt;. Proposals land in a pending area; nothing enters your memory until you confirm it. This is the spine of the whole product.&lt;/li&gt;
&lt;li&gt;Unconfirmed proposals auto-archive after 24 hours and stop cluttering the main line.&lt;/li&gt;
&lt;li&gt;Four public vaults, separated: main line / to-do / archive / reference. Search defaults to the main line; history is something you open deliberately.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The side effect: your store grows much slower than a "fully automatic memory" would. That's the point — everything in it is something you personally approved.&lt;/p&gt;

&lt;h2&gt;
  
  
  02 · Stale facts weigh the same as current ones
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;"The agent gets distracted because &lt;em&gt;every old fact has the same weight as the current one. That's a salience problem.&lt;/em&gt; What works in production for me is &lt;em&gt;typed memory with very different decay curves&lt;/em&gt;: personality and relationships are essentially permanent. Preferences fade in months. Stated intent fades in weeks. Emotion and events fade in days."&lt;/p&gt;

&lt;p&gt;— mtrifonov, Hacker News&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;strong&gt;The trap:&lt;/strong&gt; thinking this is a ranking problem. It's a &lt;strong&gt;"how long should this be kept"&lt;/strong&gt; problem. No amount of ranking rescues a memory that stopped being true.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Why it happens:&lt;/strong&gt; inverted indexes and vector similarity measure "how similar", never "does this still hold". They have no concept of time.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How we escaped:&lt;/strong&gt; &lt;strong&gt;halfway, and we'll say so plainly.&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;What we do have: separate vaults for long-lived principles vs one-off tasks; soft delete with an undo bin; the archive vault excluded from main-line search by default.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What we don't have: typed decay.&lt;/strong&gt; There is no mechanism where preferences fade over months, intent over weeks, events over days. That comment gives a workable design. We agree with it. This version doesn't implement it.&lt;/li&gt;
&lt;li&gt;It's filed as &lt;code&gt;S-20261026-02&lt;/code&gt; for the next release. &lt;strong&gt;Until then, our memories do not age on their own.&lt;/strong&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  03 · The granularity trap
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;"I still haven't found useful 'memory'. It's either an agents.md with a &lt;em&gt;high level summary, which is fairly useless for specific details&lt;/em&gt;... or something detailed... &lt;em&gt;which seems to give too much detail such that it gets ignored&lt;/em&gt;, or &lt;em&gt;detail from one functional area contaminates the intended changes in another functional area&lt;/em&gt;."&lt;/p&gt;

&lt;p&gt;— Incipient, Hacker News&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;strong&gt;The trap:&lt;/strong&gt; assuming "one memory" is a well-defined unit. It's pulled in three directions at once: summaries build context, details drive execution, and cross-project work needs isolation.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Why it happens:&lt;/strong&gt; one granularity, many purposes. A single record can't be both summary and detail, and mixing them means they contaminate each other.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How we escaped:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;One memory, one thing: &lt;strong&gt;conclusion + evidence + coordinates&lt;/strong&gt;. Coordinates mean absolute paths, line numbers, URLs — written so they can actually be checked, never "that file".&lt;/li&gt;
&lt;li&gt;Vaults carry the isolation: detail from one area never lands in another area's main line.&lt;/li&gt;
&lt;li&gt;The agent picks the vault when proposing. Pick wrong and you'll see it; fixing it costs one click.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  04 · Plaintext files bloat, and they carry noise
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;"Using MD files for this is fine &lt;em&gt;till a point&lt;/em&gt;. If you keep on adding information in your md file it will &lt;em&gt;bloat up&lt;/em&gt; and have a huge amount of data to go through, it might also have some &lt;em&gt;noise which will be picked each and every time&lt;/em&gt; that md file is read."&lt;/p&gt;

&lt;p&gt;— SachitRafa, Hacker News&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;strong&gt;The trap:&lt;/strong&gt; "plain files you can read and take with you" is our core selling point, and this lands right on it. &lt;strong&gt;Readable is not the same as clean.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Why it happens:&lt;/strong&gt; file-based storage that only ever grows is a junkyard after a year. Worse, noise doesn't disappear — every full read picks it up again.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How we escaped:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Dedupe on write; deletes are soft, into a bin you can undo (not &lt;code&gt;rm&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;Archive kept separate from the main line, excluded from default search — noise sinks, but isn't lost.&lt;/li&gt;
&lt;li&gt;Automatic daily backup, 10 copies retained, so a corrupted store can go back to yesterday.&lt;/li&gt;
&lt;li&gt;It does get slower as it grows, and that's real: measured p95 of ~127 ms at 1,356 entries; at 1M entries under load test, 0.6 s warm search but 5.3 GB of RAM. &lt;strong&gt;Plain files scale to a million entries — the price is memory.&lt;/strong&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  05 · "Memory" is actually ten different things
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;"It may seem useful at first, but eventually you'll hit a wall... &lt;em&gt;what people want by 'memory' is actually 10 different things which all need their own solution.&lt;/em&gt;"&lt;/p&gt;

&lt;p&gt;— 0xbadcafebee, Hacker News&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;strong&gt;The trap:&lt;/strong&gt; building one universal memory to cover everything. Facts, preferences, commitments, relationships, emotions, tasks — different decay curves, different notions of correctness.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Why it happens:&lt;/strong&gt; "memory" is one word in the pitch and at least six different data structures with six different freshness rules in the implementation.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How we escaped:&lt;/strong&gt; we accepted it, then &lt;strong&gt;split by purpose rather than by type&lt;/strong&gt; — main line / to-do / archive / reference, plus four encrypted private vaults.&lt;/p&gt;

&lt;p&gt;That's a trade-off. Purpose-based splitting matches human intuition ("this belongs in to-do"); type-based splitting is easier for a machine ("this is a preference, decays monthly"). We chose the former because this store is meant for &lt;strong&gt;you&lt;/strong&gt; to approve, not for a model to tune automatically. The cost is exactly pitfall 02 — typed decay still has to be built separately.&lt;/p&gt;

&lt;h2&gt;
  
  
  06 · The selling point we got wrong
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;Of 632 comments, those touching &lt;em&gt;privacy / data ownership&lt;/em&gt; numbered &lt;em&gt;4 (0.6%)&lt;/em&gt;; those touching &lt;em&gt;lock-in / portability&lt;/em&gt;, &lt;em&gt;5 (0.8%)&lt;/em&gt;. Those touching &lt;em&gt;retrieval, vectors, graphs&lt;/em&gt;: &lt;em&gt;165 (26%)&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;— our own count; method at the end of this page&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;strong&gt;The trap:&lt;/strong&gt; we had been leading with "your data stays yours, no cloud". In this corpus, &lt;strong&gt;that's barely a topic&lt;/strong&gt;. Technical audiences ask "does it work" before "who owns it".&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Why it happens:&lt;/strong&gt; data sovereignty hurts &lt;strong&gt;after a loss&lt;/strong&gt;; inaccurate memory hurts &lt;strong&gt;every day&lt;/strong&gt;. People fix the daily pain first. (Our inference, not something the comments state directly.)&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What we changed:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;For technical communities: lead with "remembers accurately, you approve it, you can undo it". Sovereignty is the foundation, not the slogan.&lt;/li&gt;
&lt;li&gt;For ordinary users, sovereignty is still the main story — their worry is "will my stuff be wiped some day".&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Two audiences cannot be sold with one sentence.&lt;/strong&gt; That was the biggest lesson here.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This one was our mistake, not someone else's pitfall. It's written down for anyone else building in this space: don't assume that because local encryption is harder to build, it's worth more in the market.&lt;/p&gt;

&lt;h2&gt;
  
  
  Data and method (checkable)
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Source:&lt;/strong&gt; Hacker News Algolia public API (&lt;code&gt;hn.algolia.com/api/v1&lt;/code&gt;), no auth. Full comment trees via &lt;code&gt;/items/&amp;lt;id&amp;gt;&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Sample:&lt;/strong&gt; 632 comments (top-level and nested) across 11 threads on AI-memory products and debates, scraped 2026-10-26. Threads were selected by keyword; scores ranged 34–86.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Topic shares:&lt;/strong&gt; keyword pre-screen, not hand-labelled per comment. So the ratios in pitfall 06 are &lt;strong&gt;order-of-magnitude&lt;/strong&gt;, not a precise statistic.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Selection bias (stated openly):&lt;/strong&gt; these 11 threads were read by people already interested in memory. They don't represent all developers, and certainly not ordinary users.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Not covered:&lt;/strong&gt; Reddit (DNS-blocked, unreachable), V2EX (site unreachable; its search API works but wasn't run). So this is a single-community sample.&lt;/li&gt;
&lt;/ul&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Why a whole page for the criticism:&lt;/strong&gt; because repeating "we're safer" long enough starts to work on yourself. Of these six, 01, 03 and 04 are places we genuinely did the work; 02 and 05 we got halfway; 06 we were wrong. Laying out the proportions honestly is more useful than claiming we solved everything — it also lets you judge whether this thing is worth installing at all.&lt;/p&gt;
&lt;/blockquote&gt;




&lt;p&gt;If you want to check any of it: the code is MIT at &lt;a href="https://github.com/330899935/apiary-memory" rel="noopener noreferrer"&gt;github.com/330899935/apiary-memory&lt;/a&gt;, the package is &lt;code&gt;apiary-memory&lt;/code&gt; on PyPI, and the benchmark script plus its raw output sit in the same repo under &lt;code&gt;/tree/main/bench&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Originally published at &lt;a href="https://hello-fengchao.com/en/pitfalls.html" rel="noopener noreferrer"&gt;hello-fengchao.com&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>opensource</category>
      <category>discuss</category>
    </item>
    <item>
      <title>Three Ways AI Memory Fails: Too Much, Nothing, Neither Trusted</title>
      <dc:creator>Feng Chao</dc:creator>
      <pubDate>Tue, 27 Oct 2026 03:47:52 +0000</pubDate>
      <link>https://dev.to/fengchao/three-ways-ai-memory-fails-too-much-nothing-neither-trusted-4gil</link>
      <guid>https://dev.to/fengchao/three-ways-ai-memory-fails-too-much-nothing-neither-trusted-4gil</guid>
      <description>&lt;p&gt;I spent two days reading what people actually say about AI memory. Not the launch posts — the comments. 632 of them across eleven Hacker News threads, plus 83 replies on a Chinese developer forum.&lt;/p&gt;

&lt;p&gt;Three sentences stuck with me. On the surface they have nothing to do with each other.&lt;/p&gt;

&lt;p&gt;They are the same problem facing three different directions.&lt;/p&gt;

&lt;h2&gt;
  
  
  The three sentences
&lt;/h2&gt;

&lt;p&gt;The first is from a Hacker News commenter called SwellJoe, replying to a Show HN for an AI memory product with biological decay:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;I know everybody seems to want the agent to remember every conversation they've ever had with it, but I just don't see the value in that. In fact, it seems to hurt productivity to have the agent second guessing me based on something I said yesterday. Every time I've used any memory system, the agent gets distracted from the current tasks based on previous conversations … often comingling unrelated projects.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;And then, further down:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;The agent is not my friend. I don't need it to remember my birthday or the nasty thing I said about React last week.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Read that again. He is not saying "I don't need memory." He is saying &lt;strong&gt;every&lt;/strong&gt; memory system he has used did this. That is not a beginner's complaint.&lt;/p&gt;

&lt;p&gt;The second is from someone on V2EX who spent two weeks building a memory system, then gave up and moved to the cloud:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Two weeks in, the system itself was fine. Every problem was about &lt;strong&gt;keeping it alive&lt;/strong&gt;.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The third, from a different V2EX thread, asked politely:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;If the AI remembers something wrong, or treats a one-off idea as a long-term preference — don't I still have to check and clean it up regularly? Compared to just maintaining Notion or Obsidian myself, how much work does this actually save?&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Three directions. One shape:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Remembering too much. Remembering nothing. Trusting neither.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I call them the three ways AI memory fails. I have hit all three, in that order.&lt;/p&gt;

&lt;h2&gt;
  
  
  Failure one: remembering too much
&lt;/h2&gt;

&lt;p&gt;Start with the counterintuitive part. &lt;strong&gt;Good memory can make a worse product.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The diagnosis came from another commenter, mtrifonov, and it is the single most useful thing in those 632 comments:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;The agent gets distracted because every old fact has the same weight as the current one. That's a salience problem.&lt;/p&gt;

&lt;p&gt;What works in production for me is typed memory with very different decay curves. Personality and relationships are essentially permanent. Preferences fade in months. Stated intent fades in weeks. Emotion and events fade in days. Reinforcement (repeated recall) keeps things alive regardless of type.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Two things there, and the second matters more.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Memory is a ranking problem, not a storage problem.&lt;/strong&gt; You have 500 facts in the store. The model can carry five into context. Which five decides whether it helps you or gets in the way.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Equality is wrong.&lt;/strong&gt; We come from databases, so we treat one row as one row — all facts equal. But memory has &lt;em&gt;age&lt;/em&gt;, &lt;em&gt;type&lt;/em&gt;, and &lt;em&gt;how many times it has been raised&lt;/em&gt;. "I said three weeks ago that this week is for launch" and "I don't eat spicy food" should never carry the same weight. The first should sink. The second should never sink.&lt;/p&gt;

&lt;p&gt;mtrifonov came back later to defend that distinction against a reply arguing decay is just eviction:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Decay-as-eviction is just LRU, fair. Type-conditional half-life is worth defending, though. … Decay everything at one rate and you're back to LRU with the problems you're calling out.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;strong&gt;The trap:&lt;/strong&gt; almost everyone's first memory system — mine included — is "store everything, retrieve on demand." Storing feels great. It is only at read time that you discover every entry is competing with today's task for a slot.&lt;/p&gt;

&lt;p&gt;We are guilty of this one. We separate &lt;em&gt;pending&lt;/em&gt; from &lt;em&gt;committed&lt;/em&gt;, but we do not weight by memory type. So yes — in our system too, the AI can drag up something from three weeks ago and get in your way today.&lt;/p&gt;

&lt;h2&gt;
  
  
  Failure two: remembering nothing
&lt;/h2&gt;

&lt;p&gt;This one stings because it is not an algorithm problem. It is an ops problem.&lt;/p&gt;

&lt;p&gt;The V2EX user hit three walls in two weeks. Keeping it alive: the whole point is to be online and absorbing, but his main machine is a laptop that sleeps when closed; he moved it to a desktop, and a system update rebooted it — &lt;strong&gt;two days before he noticed&lt;/strong&gt;. Where the data lives: memory is the entire value of the system, local disk felt flimsy, automatic backup was another pile of work. And moving machines: the archive was at home while he was at the office.&lt;/p&gt;

&lt;p&gt;His conclusion was blunt, and it is a verdict on every "local-first" pitch, including mine:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Either you accept being your own sysadmin, or you accept a subscription. There is no have-it-both-ways option.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;I want to add one qualifier, because that sentence merges two different problems.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Staying alive and being available are different architecture questions.&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;If your memory product is &lt;strong&gt;continuously ingesting&lt;/strong&gt; — eating meeting notes, email, chat — it has to run all the time. A laptop that sleeps breaks it. That constraint is real and has no clever fix. That is what he hit.&lt;/li&gt;
&lt;li&gt;If your product is &lt;strong&gt;retrieve-on-demand&lt;/strong&gt; — the model calls it when needed, it exits after — then "keep it alive" does not apply. It does not need to be alive. It needs to &lt;em&gt;be there&lt;/em&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Mine is the second kind. That is not cleverness on my part; it is a free benefit of an architecture choice. And precisely because it was free, I did not notice how much backup mattered.&lt;/p&gt;

&lt;p&gt;Here is what we actually ship: &lt;strong&gt;the entire memory is one SQLite file.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;One file. Move to a new machine, copy the folder, done. Sounds like a feature, right?&lt;/p&gt;

&lt;p&gt;It also means &lt;strong&gt;delete it and it is gone.&lt;/strong&gt; No cloud copy, no trash, no seven-day recovery. One 40 KB file holding most of what you have said for months.&lt;/p&gt;

&lt;p&gt;So the truth about the second failure is: &lt;strong&gt;"it forgot" usually means "you never backed it up."&lt;/strong&gt; Not the model's fault. The human's.&lt;/p&gt;

&lt;h2&gt;
  
  
  Failure three: trusting neither
&lt;/h2&gt;

&lt;p&gt;The hardest one, and the one nobody wants to write about.&lt;/p&gt;

&lt;p&gt;The person who asked about maintenance cost led with "I'm not a heavy programmer," taking himself out of the technical crowd. The scenario he described was familiar:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Each individual task works fine. But come back a few days later, or switch sessions, or switch agents, and the context is gone. Conclusions we already ruled out get re-argued. Writing habits I already corrected have to be corrected again.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;That is exactly the problem worth solving. Then he asked three questions, and I could not answer a single one cleanly.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;One: what if it remembers something wrong?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;It will. This is not a probability, it is a certainty. You say "let's park that approach for now" and it files "user dislikes that approach." Three months later it argues from that, and you cannot even remember saying it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Two: what if a one-off becomes a permanent preference?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;More common, more annoying. You say "skip the tests this time, we're in a hurry." It files "user does not like writing tests."&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Three, the killer: how is this different from Obsidian?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;This one hurts because at first glance there is no difference. If my memory is a pile of markdown files, why not just take notes in Obsidian and point the model at them?&lt;/p&gt;

&lt;p&gt;The only answer I have found that survives scrutiny:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Obsidian is for you to read. Memory is for the model to use. Those two things demand different standards of trust.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;When you maintain your own notes, errors get fixed and gaps get filled, because you know what you wrote. But you cannot &lt;em&gt;see&lt;/em&gt; the model's memory. There are 300 entries; you are not reading them daily. So when one is wrong, you will not find out — &lt;strong&gt;until it has spent a full day working from a false premise.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The fix for failure three is therefore not "remember more accurately." It is &lt;strong&gt;make it inspectable, make it editable, make its influence visible.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Concretely: every memory has a &lt;strong&gt;source&lt;/strong&gt;, a &lt;strong&gt;state&lt;/strong&gt;, and can be &lt;strong&gt;revoked&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;That is why ours ended up as one memory per entry with a pending queue, instead of one big file you dump into. Not for elegance. &lt;strong&gt;Memory you cannot audit is memory nobody will trust after a month.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The common shape
&lt;/h2&gt;

&lt;p&gt;All three failures are the same kind of mistake: not a technology shortfall, a trade-off made in the wrong place.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Too much&lt;/strong&gt; — the trade-off happened at &lt;em&gt;write&lt;/em&gt; time (store everything) instead of &lt;em&gt;read&lt;/em&gt; time (rank at retrieval).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Nothing&lt;/strong&gt; — the trade-off happened at the &lt;em&gt;lightweight&lt;/em&gt; end (no backup, no migration) and pushed the cost onto the user.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Trust&lt;/strong&gt; — the trade-off happened at the &lt;em&gt;automation&lt;/em&gt; end (fully automatic writes) and gave up the human's final say.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I have read a lot of technical debate about memory, and it is all vectors versus graphs, RAG versus fine-tuning. In 632 English comments and 83 Chinese replies, &lt;strong&gt;nobody was arguing about that.&lt;/strong&gt; They were arguing about the three things above.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where we stand
&lt;/h2&gt;

&lt;p&gt;I am not going to write a paragraph about how great my product is, so here is the ledger. It is called &lt;a href="https://hello-fengchao.com/?ref=devto" rel="noopener noreferrer"&gt;Apiary Memory&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What works:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Pure SQLite plus markdown. No vector store, no graph database, no daemon.&lt;/li&gt;
&lt;li&gt;Six MCP tools, so any MCP-capable client can connect.&lt;/li&gt;
&lt;li&gt;One memory per entry, with a pending queue: the model &lt;em&gt;proposes&lt;/em&gt;, you approve, then it is committed.&lt;/li&gt;
&lt;li&gt;Moving machines is copying one folder — though see gap two, because this is not written down anywhere, which means it effectively does not exist.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;What does not:&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;No type-based decay.&lt;/strong&gt; So "remembers too much and gets distracted" can only be fixed by you deleting things by hand. Filed as a work item; not started.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Backup has no documented path, only a fact.&lt;/strong&gt; The fact is that you copy &lt;code&gt;~/.memory-hive&lt;/code&gt; yourself. Nobody wrote that down, so it does not exist. Forget once and it is all gone.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;"Principle" and "one-off opinion" are still indistinguishable.&lt;/strong&gt; Same failure as number one, wearing a different hat — and the fastest way to make someone abandon the product.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;I am not going to call these "coming soon." Right now, they are missing.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The part I did not expect:&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I found these three traps because I watched other people fall into them first, then looked down and saw I was standing in one myself.&lt;/p&gt;

&lt;p&gt;Without those 632 comments I would still believe "zero dependencies, local storage, your data is yours" was a clean pitch — and I would have led with it.&lt;/p&gt;

&lt;p&gt;The reality: &lt;strong&gt;that pitch gets almost no response in this community.&lt;/strong&gt; By my own keyword count of those 632 comments, privacy and data ownership came up four times (0.6%). Lock-in came up five times (0.8%). They do not care who owns the memory. They care whether it is accurate, and whether it helps or gets in the way.&lt;/p&gt;

&lt;p&gt;That finding killed the first promotional post I had written.&lt;/p&gt;

&lt;h2&gt;
  
  
  One last thing
&lt;/h2&gt;

&lt;p&gt;After all of this, memory has almost nothing to do with database design.&lt;/p&gt;

&lt;p&gt;It is closer to a &lt;strong&gt;relationship&lt;/strong&gt;. Someone who remembers every word you ever said, and keeps bringing it up, is not thoughtful — they are unsettling. Someone who forgets everything the moment you turn around is also useless.&lt;/p&gt;

&lt;p&gt;Where that line sits, nobody has given a standard answer.&lt;/p&gt;

&lt;p&gt;But at least three failure modes can be named: &lt;strong&gt;too much, nothing, and neither trusted.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Knowing which one you are committing is worth more than knowing whether to use vectors or graphs.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Quotations from Hacker News are verbatim from the original threads (HN item 47914367 and replies, April 2026). Quotations from V2EX are translated from the Chinese; the three-wall list is condensed from one user's write-up, not a verbatim transcript.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;This piece was written in Chinese first — &lt;a href="https://juejin.cn/post/7700891797551906826" rel="noopener noreferrer"&gt;read the original on Juejin&lt;/a&gt; if you read Chinese.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>mcp</category>
      <category>discuss</category>
    </item>
    <item>
      <title>Can You Take Your AI Memory With You When You Switch Tools?</title>
      <dc:creator>Feng Chao</dc:creator>
      <pubDate>Fri, 23 Oct 2026 07:38:29 +0000</pubDate>
      <link>https://dev.to/fengchao/can-you-take-your-ai-memory-with-you-when-you-switch-tools-5doi</link>
      <guid>https://dev.to/fengchao/can-you-take-your-ai-memory-with-you-when-you-switch-tools-5doi</guid>
      <description>&lt;p&gt;&lt;strong&gt;Yes — if your memory is plain files instead of an account inside an app.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;That's the whole answer. If it's files, you can copy them, back them up, delete them, and feed the same directory to ChatGPT, Claude, Cursor, and a local model. If it only opens inside one product, it isn't yours. You just have a revocable access grant.&lt;/p&gt;




&lt;h2&gt;
  
  
  What's actually happening to multi-tool users
&lt;/h2&gt;

&lt;p&gt;If you run three or more assistants, your memory currently looks like this:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;ChatGPT knows what you do for a living and roughly how you like to be talked to&lt;/li&gt;
&lt;li&gt;Claude knows which repo you've been rewriting&lt;/li&gt;
&lt;li&gt;Cursor knows your conventions and the shape of your folder tree&lt;/li&gt;
&lt;li&gt;Whatever runs in your terminal knows something else entirely&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;None of them talk to each other. Every new window starts with the same chore: &lt;strong&gt;re-introducing yourself.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The annoying part isn't the typing. It's that models don't ask when they've forgotten — they guess. And when they guess wrong, nothing tells you. You just notice one day that the suggestions have drifted.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three structural problems with cloud memory
&lt;/h2&gt;

&lt;p&gt;This isn't a "cloud bad" argument. These are properties of the design, not bugs in any particular product.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. It's siloed.&lt;/strong&gt; Each assistant knows only its own slice. Context you gave one is invisible to the others unless you repeat it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. It's fragile.&lt;/strong&gt; Account changes, policy changes, product sunsets — months of accumulated context, gone. You had no backup, because you never received a file.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. It's rented.&lt;/strong&gt; This is the one people skip past. You think you have "my memory." What you actually have is read access to rows in someone's database. Stop paying and the access stops.&lt;/p&gt;

&lt;h2&gt;
  
  
  The only test that matters
&lt;/h2&gt;

&lt;p&gt;Ask one question:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Can I copy the whole directory to another disk and have it still work?&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Yes → it's yours. No → it's rented.&lt;/p&gt;

&lt;p&gt;Three concrete checks:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Can you open it in a text editor?&lt;/li&gt;
&lt;li&gt;Can you read it with the network unplugged?&lt;/li&gt;
&lt;li&gt;Is it still there after you stop paying?&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;All three yes, and you own it.&lt;/p&gt;

&lt;h2&gt;
  
  
  So the real fix isn't another memory app
&lt;/h2&gt;

&lt;p&gt;It's &lt;strong&gt;moving memory out of an app and back onto files.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The minimal viable shape is unglamorous — three pieces:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;A directory of Markdown or JSON files, somewhere you chose&lt;/li&gt;
&lt;li&gt;An index (SQLite is plenty; full-text search may be enough)&lt;/li&gt;
&lt;li&gt;An MCP server that exposes that directory to any MCP-capable client&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That's it. What it buys you: &lt;strong&gt;ChatGPT, Claude, Cursor, and Gemini CLI all read the same files.&lt;/strong&gt; Which client ships a breaking change, gets acquired, or shuts down stops being your problem.&lt;/p&gt;

&lt;p&gt;Memory stops being "a feature of a product" and becomes "a directory you own." Nothing about that is novel — it's how your notes and your code already work.&lt;/p&gt;

&lt;h2&gt;
  
  
  The mistake I made first
&lt;/h2&gt;

&lt;p&gt;When I set mine up, I added one rule: &lt;strong&gt;the AI may propose what to remember, but it may not write directly.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The reason was small. It had recorded "doesn't drink" as "doesn't like drinking." Barely a difference on paper — except every restaurant recommendation afterwards quietly avoided anywhere with a wine list, including a place I like. And I had no idea it had misremembered. I only knew the suggestions had gotten worse.&lt;/p&gt;

&lt;p&gt;Automatic memory's real failure mode isn't forgetting. It's &lt;strong&gt;confidently remembering the wrong thing, silently.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;So the flow became three steps:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The assistant decides something is worth keeping and writes it to a &lt;em&gt;staging&lt;/em&gt; area&lt;/li&gt;
&lt;li&gt;Staging does not touch the live store&lt;/li&gt;
&lt;li&gt;I skim it on my machine, promote what's right, drop what isn't&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Slower than full auto. But the resulting memory has been checked by me, rather than being "what a model inferred I am."&lt;/p&gt;

&lt;p&gt;If you're building your own, this is worth copying whether or not you use anyone's product: &lt;strong&gt;separate the write path from the read path.&lt;/strong&gt; Reads can be automatic. Writes probably shouldn't be.&lt;/p&gt;

&lt;h2&gt;
  
  
  What's actually available right now
&lt;/h2&gt;

&lt;p&gt;I'm not going to pretend there's only one option. Here's the current field, with the cost of each spelled out:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Option&lt;/th&gt;
&lt;th&gt;Good fit for&lt;/th&gt;
&lt;th&gt;The catch&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;basic-memory&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Plain Markdown, opens directly in Obsidian&lt;/td&gt;
&lt;td&gt;AGPL. Install needs &lt;code&gt;--prerelease&lt;/code&gt; or it silently resolves an old version&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;mcp-memory-service&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Several agents at once, REST + remote MCP, browser clients&lt;/td&gt;
&lt;td&gt;Browser access means tunneling a local port to public HTTPS — which reopens the door you were trying to close&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;@modelcontextprotocol/server-memory&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Five minutes of hands-on testing&lt;/td&gt;
&lt;td&gt;A JSON knowledge graph. No semantic retrieval&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Mem0 (self-hosted)&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;The most complete extraction and conflict handling&lt;/td&gt;
&lt;td&gt;Defaults to LLM-based extraction ⇒ your content goes back to a model vendor. Directly at odds with the goal&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Agentic Brain&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Coding-tool users; ships real benchmarks (LongMemEval-S 94.2%)&lt;/td&gt;
&lt;td&gt;Coding-agent shaped. Not a general second brain&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;SuperLocalMemory&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Local-first with an explicit "I don't know" recall contract&lt;/td&gt;
&lt;td&gt;Media indexing needs ~16 GB RAM; skips Intel Macs&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;Hive&lt;/strong&gt; (&lt;a href="https://hello-fengchao.com/?from=devto" rel="noopener noreferrer"&gt;hello-fengchao.com&lt;/a&gt;)&lt;/td&gt;
&lt;td&gt;Want a UI, and writes must be human-approved&lt;/td&gt;
&lt;td&gt;The app is macOS/iOS only — no Windows. The Python package (&lt;code&gt;apiary-memory&lt;/code&gt;) runs on its own without the app, but the database is then plain SQLite, unencrypted&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  An honest boundary: this does nothing for half of you
&lt;/h2&gt;

&lt;p&gt;One hard constraint that rarely gets stated clearly:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Local stdio MCP only works with desktop and CLI clients. Browser clients cannot connect to it.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;So if you live entirely inside chatgpt.com and claude.ai in a browser tab, none of the above helps you today.&lt;/p&gt;

&lt;p&gt;Making a browser client work requires exposing your machine behind public HTTPS (a tunnel, a reverse proxy, OAuth in front). At which point &lt;strong&gt;"data never leaves my machine" is no longer true&lt;/strong&gt; — it's just a memory sitting on a different server.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;So check this before you install anything:&lt;/strong&gt; do you have at least one MCP-capable desktop client? Claude Desktop, Cursor, VS Code, Codex, Gemini CLI — any one counts.&lt;/p&gt;

&lt;p&gt;If you don't, this whole path is closed to you for now. Wait, or paste context manually.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where to actually start
&lt;/h2&gt;

&lt;p&gt;Don't build the full stack on day one. Start here:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Take the background you keep re-typing and write it into a single &lt;code&gt;.md&lt;/code&gt; file you own.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;That's the whole first step. Two things happen: you realize how much you've been re-explaining, and you realize re-pasting the same file is tedious — at which point automation becomes obvious rather than aspirational.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Files first, tooling second.&lt;/strong&gt; Reverse that order and you end up with ten memory apps installed and none of them in use — which is the most common outcome in this category, and the reason so many of these projects have one star.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Can I export my ChatGPT memory?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Depends what you mean by export. ChatGPT will hand you a ZIP of your conversations, and since 2026 Claude and Gemini both ship an "import memory from another AI" flow that reads a structured summary you generate and paste in.&lt;/p&gt;

&lt;p&gt;So the data moves. What does not move is the part that matters: you get a profile snapshot — conclusions, preferences, stated instructions — not the conversations that produced them, and not anything the model inferred but never wrote down. It is one-way, one-time, and region-gated (the Gemini importer does not work in the EEA, UK or Switzerland).&lt;/p&gt;

&lt;p&gt;A download is not portability, and an import is not ownership.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Won't plain Markdown files be slow to search?&lt;/strong&gt;&lt;br&gt;
Under a few thousand entries, SQLite plus full-text indexing is more than enough. Retrieval speed isn't the bottleneck — &lt;strong&gt;whether you'll maintain it&lt;/strong&gt; is. Which is exactly why readability and editability beat raw performance.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How does the model know which files to read?&lt;/strong&gt;&lt;br&gt;
MCP. You expose a "recall" tool; it calls that tool when it needs context. You can also scope it — some directories readable, others never visible to a given client. The protocol doesn't define any of this, so check what each implementation actually enforces.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is this only for developers?&lt;/strong&gt;&lt;br&gt;
Configuring MCP currently is. But "keep your AI background in a file you own" isn't. The real threshold isn't technical skill — it's &lt;strong&gt;whether you have two or more assistants that need the same context.&lt;/strong&gt; With one, built-in memory is fine and this isn't worth the effort.&lt;/p&gt;




&lt;h2&gt;
  
  
  What I run
&lt;/h2&gt;

&lt;p&gt;I use &lt;strong&gt;Hive&lt;/strong&gt; — one memory server shipped in two parts. A Python package (&lt;code&gt;apiary-memory&lt;/code&gt;) on PyPI carries its own SQLite store and runs by itself on any OS. A macOS/iOS app adds the encrypted vaults, an evolution log and a UI. Six tools either way; writes still need my approval. Without the app the database is plain SQLite — not encrypted.&lt;/p&gt;

&lt;p&gt;One design choice I think more projects should copy: it publishes an &lt;code&gt;llms.txt&lt;/code&gt; at the site root, so an agent that lands on the site can read what the tools are, how to install them, and which version is required. That file is written for machines, not for a landing-page conversion.&lt;/p&gt;

&lt;p&gt;It isn't the only option. Anything in the table above gets you to "the memory is mine" — the differences are how much UI you want and whether writes require a human.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The choice matters less than the direction: start treating your memory as a directory, not an account.&lt;/strong&gt;&lt;/p&gt;




&lt;p&gt;&lt;em&gt;If the "re-explain myself to every assistant" problem is yours, the Hive integration docs are at &lt;a href="https://hello-fengchao.com/?from=devto" rel="noopener noreferrer"&gt;hello-fengchao.com&lt;/a&gt;. Or pick a different row from the table. Either way — move the files first.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>productivity</category>
      <category>opensource</category>
    </item>
  </channel>
</rss>
