# Building an AI Memory System That Doesn’t Pretend to Know Me *I exported my AI archives, built a local evidence layer, and started turning years of scattered conversations into reviewed memory. Not because I want a chatbot with a bigger backpack. Because partnership requires provenance.* For a long time, a strange amount of my working life lived inside AI chats. Some of it was junk. Some of it was repetition. Some of it was the usual model-roulette where an assistant forgets what it said three messages ago and confidently invents connective tissue where memory should be. But some of it mattered. There were article drafts. Notes about who I am and what I’m trying to build. Work on *Barely, But Here*. Professional positioning. Half-formed systems. Ideas I abandoned too early. Conversations where the AI was not just answering questions but helping me think. And all of that was trapped inside the interfaces where it happened. ChatGPT had one archive. Claude had another. Each export had its own shape, its own gaps, its own weirdness. None of it added up to something I could trust as memory. So I started building an archive. ## The archive is not the memory The local project lives at: `/Users/bertmahoney/workspace/ai-archive` That path matters less than the principle behind it. The pipeline is deliberately boring: ```text raw exports -> normalized records -> chunks -> searchable index ``` The raw exports stay immutable. They are the source of truth. I do not want a system that quietly rewrites my history in the name of convenience. From there, the archive gets normalized and indexed so I can search across it. But the index is not the brain. It is evidence. It lets me ask better questions: - What did I actually say? - What draft came from this thread? - Where did this project idea first appear? - Is this a current belief or an old note I outgrew? That distinction matters: - **Archive**: source material and evidence. - **Candidate memory**: a proposal extracted from that evidence. - **Approved memory**: reviewed material that can become part of the durable working relationship. - **Current chat context**: the temporary workbench. The archive can be large and messy. Memory should not be. That sounds obvious until you use these systems long enough. Then you realize how tempting it is to pour everything into context and call it intelligence. It is not intelligence. It is hoarding. ## What I actually want Bigger context windows are useful. I like them. I use them. But a bigger context window does not solve the underlying problem. A context window is temporary. It is where the immediate work happens. It is not a governed memory system. If I dump thousands of chunks into a model and ask it to “know me,” I get the same basic problem with more expensive failure modes. The model may retrieve the wrong thing. It may flatten contradictions. It may mistake a discarded draft for a current belief. It may treat a private note as reusable public context. That is not partnership. That is rummaging. What I want is closer to an Open Brain: a shared memory layer that supports ongoing collaboration across tools and interfaces, but only after review. Not a dump. A reviewed substrate. Something the AI and I can rely on because it was installed intentionally. ## Review before install The first governance rule is simple: personal memories are reviewed before installation. That means the archive can propose, but it cannot absorb on its own. Some candidates may be useful: - recurring facts about my work and identity - durable preferences about writing voice - ongoing projects that should survive beyond one chat - context about income, urgency, and professional positioning Some candidates may be wrong, stale, too private, or tied to a temporary emotional state. Some may be useful only with a source link attached. Some should be deleted. The review step is where the relationship becomes less creepy and more honest. It lets me say: - yes, this is still true - no, that was a draft - yes, remember this project - no, do not turn that sentence into a permanent identity claim ## The first memory scope I am not trying to memorialize everything. The first install scope is narrow on purpose: 1. *Barely, But Here* and article drafts across my ChatGPT and Claude backups. 2. My identity, professional positioning, and writing voice. 3. The AI Systems Assessment campaign, because I need income and that need is not abstract. That last point matters. A memory system for creative and professional work should remember not just preferences and projects, but pressure. Real constraints. If I am trying to build an offer that can generate near-term income, the assistant should not treat every conversation like a blank whiteboard. It should remember the campaign, the positioning, the assets, the decisions already made, and the difference between exploration and execution. But again: only because those things were reviewed and installed. Not because they appeared somewhere in a half-finished chat six months ago. ## Why this is not just a technical project I made diagrams, because eventually architecture has to leave prose. The rough shape is simple: - Hermes connects through the WebUI and other interfaces. - It can reach the local archive. - It can reach an Open Brain / reviewed-memory layer. - It can reach databases, model providers, and external APIs. The archive is not the brain. The model is not the brain. The chat window is not the brain. The useful thing is the relationship between them. The current chat is where work happens now. The archive is where I can verify what happened before. Candidate memory is where the system says, “This seems worth keeping.” Approved memory is what I have agreed can become part of the ongoing partnership. That feels more respectful than the usual assistant pattern. The usual pattern is servant memory: remember my timezone, remember my favorite formatting, remember that I like concise answers, then wait for commands. I want more partnership and less servant. That means the AI should be able to carry forward context, notice continuity, challenge me when I am contradicting myself, and help me resume work without forcing me to re-explain my life every morning. But it also means the AI should not silently decide who I am. ## The rawness is part of the point There is something uncomfortable about exporting your AI archives. You see repetition. You see how many times you asked a version of the same question. You see drafts you were excited about and never finished. You see moments where you were trying to get a machine to help you hold a life that felt too scattered to hold alone. It is tempting to clean that up before talking about it. But the mess is the material. The archive says: here is what happened. The review says: here is what still matters. The partnership says: let’s work from there. That is the part I care about. ## Where this is going The next step is not to make the archive bigger. It will get bigger on its own if I keep working. The next step is to make the review loop better: clearer candidate packets, better source links, better ways to approve, reject, revise, and scope memory. I want an AI system that can help me write, build, sell, remember, and recover threads without pretending that everything I ever typed deserves equal weight. That requires a different posture toward memory. Not “save everything.” Not “trust the model.” Not “just increase the context window.” Something more deliberate. An archive with provenance. Candidate memories with sources. Approved memories with consent. A working context that knows it is temporary. That is the shape of the system I am building. A practical AI partnership does not come from a model that can swallow more text. It comes from memory with provenance. And from the simple refusal to let convenience pretend it is truth.