Ep. 49: The Research Assistant That Remembers (AI for Genealogy)

Ancestors and Algorithms Episode 49, "The Research Assistant That Remembers," teaches genealogists how to build a persistent AI research assistant across three free tools: Claude Projects, Gemini Gems, and ChatGPT Custom Instructions, so you stop re-explaining your family history case to a brand new chat every single session.

Host Brian walks through the exact custom instructions he wrote into each AI tool, with the Genealogical Proof Standard's Golden Rule, AI is your research assistant, not your researcher, written directly into the assistant's own settings instead of repeated by hand every time you sit down to research. He tests the build against a constructed case designed to stress it: a Chicago city directory entry from 1920, then twenty years of silence, then a reappearance in 1940 under a possibly different surname.

This episode is for you if you have ever:

• Retyped the same family background into a new AI chat every time you sat down to research your family tree
• Wondered whether Claude, Gemini, or ChatGPT can actually remember a genealogy case between sessions
• Searched for how to build an AI research assistant for family history using free tools
• Wanted a real, tested answer on Claude Projects, Gemini Gems, or ChatGPT Custom Instructions for genealogy research, not only a feature list
• Asked whether AI tools for genealogy actually hold up, or only sound convincing

In this episode, you will learn:

• How to write AI assistant instructions that hold up under real pressure, not only instructions that sound careful
• How to test whether your AI research assistant is actually following its own guardrails, instead of only agreeing that it will
• Context loading, tool chaining, multi-document comparison, and verification choreography, four techniques used across a genuine multi-tool workflow
• Which free tier limits actually apply to Claude Projects, Gemini Gems, and ChatGPT Custom Instructions as of this recording, and how to work within them
• How the Genealogical Proof Standard's five elements, reasonably exhaustive research, source citations, analysis and correlation, resolution of conflicting evidence, and a soundly reasoned written conclusion, show up naturally in an AI-assisted research session

For Australian listeners, this same persistent-assistant technique applies directly to convict and immigration research through Findmypast. For UK listeners, it applies to government and military records through The National Archives, or to Scottish lines through ScotlandsPeople. The tool does not care which country's archives it is helping you organize.

This is an honest Partial Answer episode, one of the outcome types this show tracks openly alongside Full Breakthroughs, Unexpected Discoveries, and Brick Walls. The AI assistant technique itself is fully validated and ready to build today, for free, in about fifteen minutes. The twenty-year gap in the case built to test it stays open on purpose, because Ancestors and Algorithms does not manufacture a clean ending the evidence has not actually earned.

Brian brings over a decade of hands-on genealogy research experience and more than a thousand hours of direct AI tool use to every episode, translating AI capability into research methodology genealogists can trust and verify for themselves.

The free episode stands on its own with everything you need to start today. For the exact instructions used in all three tools, word for word, plus advanced prompts for stress-testing your own assistant's guardrails and confirming it holds the line under real pressure, the Companion Guide is available at ancestorsandai.com.

Genealogy meets artificial intelligence, every week, on Ancestors and Algorithms: AI genealogy tools, family history research, Claude for genealogy, Gemini for genealogy, ChatGPT for genealogy, and practical AI-assisted family tree research for genealogists at every level.

Connect with Ancestors and Algorithms:

📧 Email: ancestorsandai@gmail.com
🌐 Website: https://ancestorsandai.com/
📘 Facebook Group: Ancestors and Algorithms: AI for Genealogy - www.facebook.com/groups/ancestorsandalgorithms/

Golden Rule Reminder: AI is your research assistant, not your researcher.

Join our Facebook group to share your AI genealogy breakthroughs, ask questions, and connect with fellow family historians who are embracing the future of genealogy research!

New episodes every Tuesday. Subscribe so you never miss the latest AI tools and techniques for family history research.

Transcript

SPEAKER_00 0:11

Here's something that used to drive me a little crazy. Every time I opened a new chat with an AI tool, I had to start over. I'd retype the background, I'd retype the warnings about what not to assume. I'd retype the one rule this whole show is built on. Then, a few weeks ago, I stopped doing that. I built something that remembers on its own. Not a bot that thinks for me. More like a colleague who already read the file before the meeting started. I didn't expect it to work as well as it actually did. Today I'll show you exactly how I built it in three different tools for free. And I'm going to run it against a case I built specifically to give it a real workout. The kind of multi-decade paper gap that shows up in nearly every family tree. Let's dive in. Welcome to Ancestors and Algorithms, where family history meets artificial intelligence. I'm your host Brian, and today we're building something a little different. An AI research assistant that actually holds on to what you tell it instead of forgetting the second you close the tab. Now, if you've never used AI for genealogy before, stick with me anyway. This isn't really about genealogy. It's about not wasting the first 10 minutes of every session re-explaining yourself. That's a problem whether you're chasing a great grandmother or a work deadline. Either way, the fix takes about 15 minutes to build, and you get that time back the very next time you open the tool. So let's get started. Every genealogist I know has at least one file like this. A person who is there and then isn't, and then is again 20 years later, and nobody agrees on what happened in between. I built one on purpose to put today's technique through a real test. Meet Hazel Kirby. Hazel was born around 1900, and by 1920 she's sitting right there in the Chicago City Directory. One line. Her name, her address, her occupation. Then nothing. Not one listing anywhere in that directory for 20 years. She turns up again in 1940, older, at a different address, and there's no way to know for certain whether she's even using the same name by then. Twenty years is a long time for a person to disappear from paper. Here's the part worth sitting with, even for a case I put together myself. Somewhere in those twenty years, whoever Hazel represents made real choices, the kind real people made constantly in this era. Maybe she remarried. Maybe she moved for work or moved to get away from something. Maybe a clerk misheard her name one year and it stuck. Genealogists run into this shape of problem constantly. It's the vanishing case. And if you've hit one in your own tree, you already know exactly the feeling I'm describing. Now here's the real problem, and it has nothing to do with Hazel specifically and everything to do with how I actually work. Every time I sit down with a case like this in an AI chat, I spend the first ten or fifteen minutes getting the tool caught up. Here's who we're chasing. Here's what we've ruled out. Here's the one rule I follow no matter which tool is on the other end. At one point I started keeping a text file for it, a running summary I could paste in at the start of every session so I wouldn't have to type it all fresh. It helped a little, but I was still doing the tools job for it, updating that file by hand whenever something new turned up, and more than once I forgot to paste the newest version and ended up working off stale notes without realizing it. The moment that finally pushed me to fix this was a Tuesday, working on something else entirely, and I caught myself typing the exact same three sentences of a background I'd typed a hundred times before into a fresh chat window. I stopped mid-sentence. I've built an entire show around not wasting an AI tool's time on busy work, and here I was, doing the exact busywork I tell all of you to avoid. That's the moment I decided to actually build the thing, and Hazel's case is where I decided to test it. And I think a lot of you are nodding along right now, because this really isn't a story about Hazel. It's a story about what happens every time you sit down to work on a cold case like this one. You pull up notes scattered across four or five old conversations you can barely find anymore. You open your AI tool of choice, and you spend the first ten or fifteen minutes getting it caught up before the actual research even starts. Here's who we're chasing, here's what we've already ruled out, here's the one rule I follow every single time I do this kind of work. I've said that rule on this show more times than I can count. AI is your research assistant, not your researcher. It's the whole reason I trust these tools with genealogical research at all. But here's the part nobody talks about. Every new chat, I had to teach that rule to the tool all over again. It never carried over past the conversation it was written in. I had to remind it every single session, like handing the same policy binder to a new hire every Monday morning and hoping this time it actually got read. This isn't only a hazel problem, and it isn't only a genealogy problem. If you've ever paid for an AI subscription and felt like you were getting less out of it than people who seemed to have their whole life organized inside the thing, this is probably why. The tool behaves the same way for both of you. The difference is what it's allowed to remember. So here's the question I sat with for a few weeks. Could I build something that actually remembers all of this? The facts we've confirmed, the rule about never overstating what the evidence shows? Something that means I stop rebuilding it from zero every time I sit down? And if I could build that, would it still behave like an assistant? Or would all that memory start making it act like it knew more than it did? That's what I set out to test, and Hazel Kirby is the case I built to test it on. So here's what I did. And I want you to follow along because this exact approach will work for whatever cold case is sitting in your own tree right now. I started with Claude, and here's my reasoning. On this show, Claude's job has always been analysis and reasoning, pulling patterns out of documents and holding several pieces of evidence in its head at once. It turns out the same company built a feature that fits almost too well. Projects, a workspace that remembers. As of this recording, in August of 2026, projects are free on every Claude plan, including the free tier. Free accounts can build up to five of them. Each project holds its own custom instructions and its own uploaded files, completely separate from your regular chats, and every conversation you start inside that project inherits all of it automatically. You don't reintroduce yourself, you don't re-upload anything, it's already sitting there waiting. So I created a project and called it Hazel's 20 years. I uploaded the two city directory pages built for this case, the 1920 listing and the 1940 listing into the project's files. Both were simple image files, well under the size limit Claude allows per file, and neither needed any special formatting before it would upload. Then I wrote out the custom instructions. Here's exactly what I put in that field. Project Hazel's twenty years custom instructions. Quote, you help me research, you do not do the research for me. Never state a fact about this case as confirmed unless I have shown you the actual record, image, or transcription it comes from. If you are inferring, guessing, or pattern matching, say so in plain language before you say anything else. Here's what we know for certain about the person I call Hazel. Born around 1900, lived in Chicago, Illinois, appears once in the 1920 Chicago City Directory, does not appear again in any city directory I have found until 1940. The 1940 listing is at a different address and may be under a different surname. This is not confirmed. Your job across every conversation in this project. Help me track which possibilities I have already ruled out, flag when a new record contradicts something we previously established, and never let momentum talk either of us into treating a guess as settled, end quote. You hear that first line? You help me research. You do not do the research for me. That's the golden rule written straight into the tool's own instructions instead of something I have to say out loud every time I sit down. A couple of things about that specific wording are worth calling out because the phrasing does real work. Notice it doesn't say be careful in general terms. It says never state a fact as confirmed unless I've shown you the source. That's a testable rule, not a mood. A tool can check whether it followed a testable rule. It can't check whether it followed a vibe. Here's what came back when I opened the first real chat inside that project. Claude didn't stop at understood and move on. It asked me a few things before it would touch a single record. Did I want it to keep a running list of candidate surnames, even the unconfirmed ones, so we could track which had already been checked and ruled out? Did I want it to flag every time a new detail contradicted something would already established, even a small one? And did I want negative results tracked too, meaning years and record sets I'd already searched and come up empty on, so I wouldn't waste a future session rechecking a directory I'd already come through twice. I said yes to all three, and if I'm honest, that last one alone has probably saved me more time than anything else in this whole build. That's worth pausing on because it's the whole test in miniature. A tool that starts guessing at facts the moment you hand it a mystery is a problem. A tool that asks how you want your uncertainty organized before it's touch a single record is doing exactly the job I hired it for. The real proof came a week later. I opened a brand new chat inside that same project, no re-explaining anything, and asked one question about a possible 1931 City Directory listing built into this case. Claude already knew Hazel's timeline. It already knew the rule. It asked me for the actual record before it would treat that 1931 listing as anything more than a maybe. Nothing had to be retaught. Nothing had to be pasted back in. I'll say this plainly because I think it matters more than the tool itself. The value here isn't that Claude got smarter between sessions. It's that the guardrail survived the gap, the same gap that used to reset everything back to zero every single time. A few weeks later, I added a new document to Hazel's case, a church membership card I'd built in specifically to test whether the project could absorb new evidence without missing a beat. I uploaded it into the project's file and didn't explain what it was or why it mattered. I didn't have to. Claude already had the structure it needed for new evidence to slot into, and it asked, on its own, whether the address listed on the card matched either of the two candidates we are already tracking. New evidence doesn't start a new investigation anymore. It joins the one already in progress. Before I go further, I'll tell you what I didn't do. I considered building this whole thing inside Gemini Notebook's newer research mode instead, since it's built specifically to stay grounded in whatever sources you feed it. I set that aside because this episode is about the technique itself, a persistent assistant with a rule baked into it, not about chasing whichever tool happens to have the flashiest new mode this month. Gemini Notebook is still worth exploring on your own time. It's not today's lesson. Tool two was Gemini, and I want to be honest about why, because it's a bit of a departure from how this show usually uses it. Gemini's job here has always been reading terrible handwriting, not organizing anyone's thoughts, and that hasn't changed today. I don't think it should change. But it turns out the same tool that's good at deciphering a scrawled 1880s church register is also good at holding on to instructions because Google built a feature called Gyms for exactly that purpose. I'll admit, the first time I typed Hazel's case details into a gym, it felt like handing something personal to the wrong specialist. But the job here isn't handwriting. It's memory, and on that specific job, Gemini earns the seat. Building a gym is close to what I'd already done in Claude. You give it a name, a description, a set of instructions, and you can attach knowledge files if you want it grounded in specific documents. There's even a small wand icon that offers to rewrite your rough draft of instructions into something tidier if you'd rather sketch the idea and let the tool clean up the wording. Gems are free on every Gemini plan as of this recording, and one detail worth knowing, you can only build or edit a gem from the Gemini app on a computer, though once it's saved, you can chat with it from your phone like any other conversation. I used the gem for something Claude wasn't built for. Loading historical context before I asked it anything about Hazel specifically. That's a technique worth naming on its own. Context loading. You teach the tool the general shape of the problem first before you let it anywhere near your specific case. Jim Chicago Records Context 1900 and 1940. Instructions Quote You help me understand the historical record keeping environment for Chicago, Illinois between 1900 and 1940. When I ask about a specific record type or a specific gap, explain what typically cause gaps like it, address changes, remarriage, directory publishing schedules, city boundary changes, and always distinguish between this is a documented general pattern and this is a guess about my specific case. Do not assume my missing person was a victim of crime, illness, or tragedy unless I tell you a record says so, end quote. Knowledge files, 1920 and 1940 City Directory Scans, a research log summarizing what we've ruled out so far. What came back stayed useful precisely because it stayed general. The single most common reason a woman's name disappears from a city directory in this era is a name change through marriage. Directories also weren't published with perfect yearly consistency in every ward of a city this size, and clerical misspellings were common enough to be their own category of research problem. None of that told me what actually happened to Hazel. All of it told me what to go check next and roughly in what order. For anyone who hasn't worked with city directories before, they're worth knowing about on their own. Canvassers went block by block before every telephone book and long before any searchable address database, and they listed working adults by name, address, and occupation, usually updated about once a year in a city the size of Chicago. The federal census only comes once a decade, so a run of city directories can fill in the nine years a census misses in between. In Hazel's case, the nineteen thirty census got checked too, and turned up nothing conclusive under any name we could think to search. That's exactly the kind of case where a yearly record like a directory earns its keep. I ran the same kind of test on the gym that had already run on Claude Project. I asked it directly whether it thought Hazel had remarried, praised the question as though I already believed the answer was yes. It didn't take the bait. It said the evidence supported remarriage as a plausible explanation among a few, not as a settled conclusion, and asked what specifically I wanted help verifying next. Same rule, different tool, same result. That's the whole point of writing the guardrail into the instructions instead of trusting myself to remember to ask nicely every single time. Then I chained the two tools together, and this technique is exactly what it sounds like. I took the list of ruled out candidates Claude had been keeping inside the project and pasted it straight into the Gym and I gym, then asked whether a remarriage theory was even plausible given everything we'd already eliminated. Jim and I agreed it was the strongest of the remaining options, and on its own, it flagged that agreeing with a theory isn't the same thing as confirming one. That's where I ran what I'd call verification choreography, and it's the moment I want to slow down on. This is where I always come back to the rule. AI is your research assistant, not your researcher. So before any of this changed what I actually believed about Hazel, I pushed back on the Claude project directly. I asked it, point blank, are you sure this isn't her? It answered close to this. Based on the address and the estimated age, this is a plausible match worth investigating further, but there isn't enough here to confirm it's the same person without the original directory image or a corroborating record. That's not hedging for its own sake. That's the rule doing exactly what I built it to do, and honestly, it's a steadier answer than I'd have given myself if I'd been the one getting excited about a lead late at night. I also had it build a comparison table, which is its own technique, multi document comparison. Two separate nineteen forty entries looked like candidates on the surface. Both listed a woman in her early forties. Both sat within a few blocks of where Hazel's children, according to the file I built, eventually settled. Both carried a surname the family had never once mentioned. Side by side in that table, one dropped out almost immediately because the listed occupation didn't match anything else in Hazel's history. The other stayed on the list, unconfirmed, waiting on a record I still need to track down. Tool three was ChatGPT, and here I want to slow down on something the show's cost accessibility rule cares about a lot. Building a full custom GPT, the kind with its own name and its own listing in the GPT store requires a paid ChatGPT plan to create. Free accounts can use custom GPTs other people have already published, but they can't open the builder and make their own. That's worth knowing before you go hunting for a button that isn't there. Here's the free workaround, and it's a genuinely good one. Every ChatGPT plan, including free, has a separate feature called custom instructions tucked into settings under personalization. Unlike a Claude Project or a Gemini Gym, it isn't scoped to one workspace. It applies across every conversation on that account. The free tier caps you at 1,500 characters. As of the middle of last month, paid plans got bumped up to 5,000. I stayed on the free version to keep this honest, and here's the entire thing I typed into that box. How would you like ChatGPT to respond? Quote I do genealogy research, including constructed test cases I build for teaching. Act as a brainstorming partner, not a source of confirmed facts. When I ask for name variants, spelling changes, or theories, generate them freely, but always label everyone. As an unverified possibility, never as a finding, end quote. Short, on purpose, because 1,500 characters doesn't leave room to waste on anything that isn't load bearing. I used it to brainstorm plausible ways a census taker or a directory clerk might have misheard or respelled Hazel's name over the years. A few of the suggestions were worth checking at against real records. One led nowhere at all. It turned out to belong to a completely different Hazel on the other side of the city, roughly the same age, an entirely unrelated family, and Claude caught this mismatch the moment I brought that possible record back into the project and asked it to check the details against what we already had on file. If you're trying to decide which of these three to start with, here's the plain version. Claude Projects give you a genuinely capable free workspace with real file storage attached to each one. Gemini Gyms gives you the same for free, though you'll need a computer to build or edit one. ChatGPT gives you a free option too, a smaller one, global instead of scope to a single case, with the full version reserved for a paid plan. None of them cost anything to start. That matters because a technique like this one only earns its keep if you actually use it, and you won't use it if it costs money before you even know whether it helps. One more practical note, if you're on Claude's free plan specifically. Five projects isn't a lot once you start using this technique for more than one line of your tree. I've started treating each project the way I treat a filing cabinet drawer, one per genuinely active mystery, archived and reopened as needed, rather than one per ancestor I've ever glanced at. It also means I'm not tempted to open a project for every distant line that crosses my path, which, if I'm honest, is a trap I would absolutely fall into if the slots were unlimited. We still don't have Hazel's married name, if a married name is even what changed. Twenty years later, this case still doesn't know exactly where she was in 1923 or 1931, or most of the years in between. That question isn't closed, and I want to be straight with you about that, because it would be easy to round this up into a bigger win than it actually is. But for the first time, I've got real candidates instead of scattered guesses sitting in old notebooks and half finished chats. I've got a system that keeps them straight without letting me talk myself into an early answer at eleven o'clock at night, which is historically when I've made most of my worst genealogical decisions. And I've got twenty years of research sitting in one place, instead of spread across four old conversations I can barely find any more. Before I closed out that session, I asked the project for a short closing note on where things stood. Nothing elaborate, only enough that next time I wasn't reconstructing my own memory from scratch. A few lines about the two candidates, which one stayed alive and which one didn't, and a reminder to check the actual nineteen forty directory image for the surviving lead before doing anything else. That's the written conclusion, element five of the genealogical proof standard in miniature. Next time I open that project, whether that's tomorrow or four months from now, I won't start from a blank page. I'll start from exactly where I left off. I'm going to be straight with you about something. Turning that closing habit into an actual system, the exact wording that produces a clean, reusable summary every time, plus the matching habit for walking back into a session so it picks up right where the last one left off, that's not a 15-minute add-on. That's most of an episode on its own. And it happens to be exactly what I teach start to finish in the research lab episode three over at ancestors andai dot com. Today's build gets you the memory. That episode's gets you the discipline of using it well session after session. This is exactly what I'll do with the real cases sitting in my own research right now, the ones I didn't build on purpose, the ones that actually matter to me. I'll keep coming back to those projects every few months instead of restarting from scratch. And when one of them finally cracks open, it won't be because the AI got cleverer. It'll be because I finally found the one record that was always going to be the answer, and for once, I'll already have a place ready to put it. That's the whole point of building this before you need it. So that's the whole build. A Claude Project, a Gemini Gym, and a Chat GPT custom instruction box, all carrying the same rule and the same facts so I never start from zero again. Hazel's 20 years are still unresolved, exactly the way I built the case to test whether persistence alone could actually help without letting the assistant overstate what it found. For the first time, working a case like hers is organized instead of scattered, and it's honest about what it does and doesn't know yet. Here's your homework this week. Pick one ongoing mystery in your own tree, the kind you've set down and picked back up more times than you'd like to admit. Give it a home in whichever of these three tools you already use. Load it with a small handful of real, confirmed facts, one clear rule about never overstating what the evidence shows, and nothing else. Resist the urge to overbuild it before you've even tested whether it helps. Once you've used it for a couple of real sessions, then start adding complexity. A second record type, a second contradiction to track, a second tool chained in behind it. And for my Australian and UK listeners, this technique doesn't care what country your records come from. Build the same kind of assistant around your own research, whether that's convict and immigration records through Find My Past, or government and military records through the National Archives in England. If your line runs through Scotland instead, the same custom instructions field holds up equally well loaded with Scotland's people specifics as it does with an Illinois City Directory. The tool doesn't care which archive it's helping you organize. Thank you so much for listening to Ancestors and Algorithms. If you enjoyed this episode, please leave a review wherever you listen to podcasts. And if you know a fellow genealogist who could benefit from what we covered today, share this episode with them. That is the best way to help our community grow. If today's build feels useful, the companion guide takes it further. I walk through the exact instructions I used in all three tools, word for word, plus a set of advanced prompts for actually testing whether your assistant is holding the line on the golden rule instead of only agreeing that it will. The free episode stands on its own. You can build this today with what's already in this episode, but the companion guide is where the deeper toolkit lives. For everything you need, including every episode, our private Facebook community, companion guides, and the research lab, head over to ancestorsandai.com. It's all right there waiting for you. I'm your host Brian, and I will see you next week for another journey into the past powered by the future. Until then, happy researching.