<?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"><channel><title><![CDATA[Mnemose]]></title><description><![CDATA[Mnemose]]></description><link>https://mnemose.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Wed, 18 Nov 2026 14:42:53 GMT</lastBuildDate><atom:link href="https://mnemose.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[dasf]]></title><description><![CDATA[What if your AI assistant could remember not just what you said, but how your life has changed since then? We've all had that moment with an AI assistant…]]></description><link>https://mnemose.hashnode.dev/dasf</link><guid isPermaLink="true">https://mnemose.hashnode.dev/dasf</guid><dc:creator><![CDATA[Anushay Jain]]></dc:creator><pubDate>Sun, 18 Oct 2026 18:23:36 GMT</pubDate><content:encoded><![CDATA[<h1>Building Mnemos: Giving AI a Memory That Actually Evolves</h1>
<p><em>What if your AI assistant could remember not just what you said, but how your life has changed since then?</em></p>
<p>We've all had that moment with an AI assistant where we have to explain something for the third time. It might be a project we're working on, a preference we mentioned last week, or a decision we made in an earlier conversation.</p>
<p>AI is getting really good at answering questions. But remembering the context behind those questions, across days, conversations, and changing circumstances, is a different problem entirely.</p>
<p>That was the idea behind <strong>Mnemos</strong>, the project I built for the <a href="https://www.wemakedevs.org/hackathons/falkordb">FalkorDB Hackathon</a>.</p>
<p>The goal was to build a local-first AI desktop workspace that captures your digital activity and turns it into memory you can actually use later. Not just a chatbot with a longer context window, but a system that can connect past experiences, track changing facts, and retrieve the right information when you need it.</p>
<p>Here's how I approached it.</p>
<h2>The problem with AI memory</h2>
<p>Most of us don't remember things as isolated pieces of text. We remember connections.</p>
<p>You might tell your assistant that you're working on a project, discuss a deadline during a meeting, and then change that deadline in a conversation a few days later. All three events are related, even though they happened at different times and in different contexts.</p>
<p>Now imagine asking your assistant, "What's the deadline for my project?"</p>
<p>A useful assistant shouldn't just retrieve the first mention of a deadline. It should understand the sequence of events, identify the latest information, and know whether an earlier statement is still valid.</p>
<p>This is where simple chat history starts to fall short. Storing more text doesn't automatically give an AI a structured understanding of how things relate or change over time.</p>
<p>I wanted Mnemos to handle that problem differently.</p>
<h2>Meet Mnemos</h2>
<p>Mnemos is a local-first AI desktop application built around the idea of persistent, connected memory.</p>
<p>It can capture screen activity, extract text through OCR, process meeting audio and transcripts, and retain chat conversations. It also supports searchable history, AI chat with voice interaction, meeting summaries, action items, and a visual interface for exploring memory.</p>
<p>The interesting part isn't just collecting all this information. It's what happens after it's captured.</p>
<p>Instead of treating everything as a pile of text, Mnemos builds a connected memory graph using <strong>FalkorDB</strong>. This graph becomes the foundation for retrieving information, understanding relationships, and keeping track of what has changed.</p>
<h2>Why FalkorDB became the heart of Mnemos</h2>
<p>The core technical decision was to use FalkorDB as the memory layer.</p>
<p>Mnemos stores captured data and searchable history in SQLite, while FalkorDB maintains a graph of experiences, facts, and conversations. Both serve different purposes, and the graph is what makes the memory-aware features work.</p>
<p>I modelled the memory system around three main building blocks:</p>
<ul>
<li><p><strong>Episodes:</strong> Individual pieces of experience, such as screen activity, meeting transcripts, notes, or chat turns.</p>
</li>
<li><p><strong>Facts:</strong> Information extracted from those experiences that should remain available over time.</p>
</li>
<li><p><strong>Sessions:</strong> Conversations that group related exchanges together.</p>
</li>
</ul>
<p>These aren't isolated records. They're connected through relationships that capture chronology, provenance, and changes in information.</p>
<p>For example, an episode can lead to a fact, that fact can belong to a conversation session, and a later fact can supersede an earlier one.</p>
<p>This is where a graph database becomes particularly useful. Instead of trying to reconstruct relationships every time someone asks a question, Mnemos can query the relationships that already exist.</p>
<h2>The tricky part: remembering when something changes</h2>
<p>This was one of the most important design decisions in Mnemos.</p>
<p>Suppose I tell my assistant:</p>
<blockquote>
<p>I live in New York.</p>
</blockquote>
<p>A week later, I say:</p>
<blockquote>
<p>I moved to Lisbon.</p>
</blockquote>
<p>A basic memory system might save both statements and leave the AI to figure out which one is correct. A similarity-based search might even retrieve the New York statement because it happens to be relevant to the question, "Where do I live?"</p>
<p>But relevance isn't the same thing as truth.</p>
<p>Mnemos treats these as two statements about the same underlying attribute: my current location.</p>
<p>When the new fact arrives, the older fact isn't simply overwritten or deleted. Instead, Mnemos marks the older fact as no longer current and creates a <code>SUPERSEDES</code> relationship from the new fact to the old one.</p>
<p>Each fact also tracks <code>current</code>, <code>valid_from</code>, and <code>valid_to</code>, allowing the system to distinguish what is true now from what was true previously.</p>
<p>That means the graph can preserve both pieces of information without confusing them.</p>
<p>Ask Mnemos where I live, and it can retrieve Lisbon as the current answer. Ask what changed, and the older New York fact is still there to provide context.</p>
<p><strong>The important distinction is that Mnemos doesn't just remember what was said. It keeps track of whether that information is still valid.</strong></p>
<h2>Why matching facts by meaning matters</h2>
<p>There's another subtle problem here: two statements about the same thing might share almost no words.</p>
<p>"I live in New York" and "I moved to Lisbon" don't look particularly similar as strings of text. Yet they clearly describe changes to the same attribute.</p>
<p>Mnemos addresses this by keying facts around the attribute they describe, rather than relying only on the wording of each statement.</p>
<p>That gives the system a way to identify conflicting claims even when their phrasing is completely different.</p>
<p>It's a small implementation detail with a significant impact. Without it, a system can accumulate contradictory memories and struggle to determine which one to trust.</p>
<h2>Retrieving memories with Cypher</h2>
<p>Once information is stored in the graph, the next challenge is retrieving it correctly.</p>
<p>Mnemos uses OpenCypher queries to retrieve relevant context from FalkorDB. The retrieval process prioritizes current facts, then considers superseded facts while identifying them as replaced, followed by related episodes that provide additional context.</p>
<p>This ordering matters because not every relevant memory should be treated equally.</p>
<p>A current fact can answer a question directly. An older fact can explain how the answer changed. A related episode can provide the original conversation or experience behind the claim.</p>
<p>Together, these give the language model more useful context than a collection of loosely related text snippets.</p>
<p>It also makes memory more inspectable. The Brain view lets you explore the underlying graph instead of treating the assistant's memory as an invisible black box.</p>
<h2>Teaching the assistant when not to answer</h2>
<p>One thing I didn't want Mnemos to do was invent personal history to fill gaps in its memory.</p>
<p>If someone asks a question and the graph doesn't contain a matching fact or relevant episode, the assistant shouldn't confidently manufacture an answer.</p>
<p>To handle this, Mnemos injects an explicit <code>[abstain]</code> instruction when retrieval finds no matching memory. The model is instructed to acknowledge that it doesn't know rather than make up a past event or personal detail.</p>
<p>This doesn't magically eliminate every possible hallucination, but it gives the model a clearer instruction about what to do when the retrieved evidence is missing.</p>
<p>For a memory system, that distinction is important. An honest "I don't know" is far more useful than a fabricated recollection.</p>
<h2>Local-first, by design</h2>
<p>Mnemos is designed to run on your machine, with its desktop application, capture API, SQLite storage, and FalkorDB instance running locally.</p>
<p>FalkorDB runs in Docker and starts alongside the application. The result is a memory system you can run without relying on a separately hosted graph database.</p>
<p>There's also a deliberate separation between capture and graph-based memory. If FalkorDB is unavailable, the capture pipeline can still write to SQLite. However, the Brain view, profile, and graph-aware chat context depend on the graph and won't have their normal memory functionality.</p>
<p>That separation makes the architecture easier to reason about: capturing activity is one responsibility, while building meaningful, connected memory is another.</p>
<p>When a Gemini API key is configured, Mnemos uses Gemini for chat and other AI-powered features. Retrieved graph context can be sent to the model to help generate responses, so local storage doesn't mean every part of the AI processing is necessarily offline.</p>
<h2>Putting it all together</h2>
<p>At a high level, the architecture looks like this:</p>
<ol>
<li><p>Mnemos captures screen activity, text, meeting audio, and conversations.</p>
</li>
<li><p>Captured content is stored locally, with SQLite providing searchable history.</p>
</li>
<li><p>Meaningful facts and experiences are written into FalkorDB as connected graph records.</p>
</li>
<li><p>New facts can supersede old ones while preserving their history.</p>
</li>
<li><p>When a question comes in, Cypher queries retrieve current facts, relevant historical claims, and related episodes.</p>
</li>
<li><p>Gemini uses that context to generate a response, with explicit abstention instructions when the memory lookup comes up empty.</p>
</li>
</ol>
<p>The result is a desktop AI workspace where memory isn't just a transcript of everything that happened. It's a structured representation of experiences and the information learned from them.</p>
<h2>What I learned building it</h2>
<p>The biggest takeaway for me was that <strong>AI memory is not simply a storage problem. It's a data modelling problem.</strong></p>
<p>You can store thousands of conversations, but that doesn't mean you know which facts are current, how decisions evolved, or where a piece of information came from.</p>
<p>Building Mnemos pushed me to think about memory in terms of relationships, history, provenance, and retrieval rather than just prompt length or text similarity.</p>
<p>It also made the choice of database much more concrete. FalkorDB wasn't added just to satisfy a technology requirement for a hackathon. The graph model directly supports the problem the project is trying to solve.</p>
<p>There is still plenty to improve, especially around the reliability of fact extraction, conflicting claims, and evaluating whether retrieved memories are genuinely useful. But this project gave me a foundation for exploring those questions in a working desktop application.</p>
<h2>What's next?</h2>
<p>I'd like to keep pushing Mnemos toward an assistant that can follow work across time without requiring users to repeatedly explain their context.</p>
<p>That means making memory updates more reliable, improving retrieval when a question connects several past experiences, and making the graph easier to inspect and understand.</p>
<p>The larger idea is simple: an AI assistant should be able to build context from your interactions instead of starting from scratch every time.</p>
<p>If you're interested in the implementation, the code is available here:</p>
<p><a href="https://github.com/Punay04/mnemos"><strong>GitHub: Punay04/mnemos</strong></a></p>
<p>Mnemos was built for the <a href="https://www.wemakedevs.org/hackathons/falkordb">WeMakeDevs FalkorDB Hackathon</a>, using FalkorDB as its core memory graph.</p>
<p>The question I'm exploring with this project is one I find increasingly interesting: what would AI assistants feel like if they could remember not just our words, but the evolving context behind them?</p>
<p>That's what I'm building toward with Mnemos.</p>
]]></content:encoded></item></channel></rss>