The most code I have ever written by hand is about ten lines of Python, and I gave up on those in 2018. Today I run Total Recall, a memory layer for AI agents that I designed and built, and it sits behind every agent I work with, remembering what we did, what I decided and why. Tom Granot spent 15 to 18 years as an engineer living in the terminal, and Total Recall became a real part of his own memory layer before he even sat down to review it on his channel, SyntaxGTM.
I don't think my story is a fluke. People who live with a problem every day, and know their field better than anyone, can now build the fix themselves, all the way down at the infrastructure level. Memory was my problem, so memory is where I started.
Big confession: I am not a developer
I'm a marketer by trade, with 15 years in the job, and my whole coding career before AI fits in one sentence.
I think the most lines of code I have ever written was about ten lines of Python, when I attempted to learn Python somewhere around 2018, and I gave up.
About a year before this conversation I was VP Marketing at a startup working on AI and data, watching the technology move so fast that following it from the side stopped feeling like an option. So I said, you know what, the hell with it, and quit my job. I didn't want to be behind, and I knew I had to learn this thing for real.
I started building my own MCPs, my own tools and agents, you name it. I also started a community in Israel for marketers working with AI, and began working with companies on AI transformation and implementation, the work I now do at 10x Company with my partners Noy Rubinstein and Hadar Yakir. Almost all of it, the client work included, ran through AI agents, which is how I ran straight into their biggest weakness.
I got tired of re-explaining myself
Agents forget. Every new session starts from zero, so the agent doesn't know which client you mean, what you tried last time or why you chose one approach over another, and you end up explaining it all again, then again the next day. We always joke that AI gives us all a kind of ADHD, and I already had that feeling, with so many projects running at once and so much to keep track of. Re-explaining myself to my own tools on top of all that wore me out.
What I wanted was a partner. I wanted to open a blank chat and say, "Hey, remember what we did with that client, with that workshop? Can you do that thing, but for this other client?" and have it just know. A partner like that changes what I can hand over, too. Once an agent remembers the context, it can run longer tasks and more intricate loops, and I only step in when my opinion is truly needed.
Tom named the missing piece precisely. Every session holds two kinds of material, what the AI produced and the human input you gave it along the way, meaning your instructions, your reasoning and your decisions. In his view, the question Total Recall answers is what you told the AI while you worked, and that human input is exactly the part you end up repeating. When Tom opened a new folder and told his agent what he wanted to build, Total Recall went and found everything he had said in earlier sessions about similar work, and used it to scaffold a plan for the next round. I get the same feeling every day, and the best part is that it happens without being asked. I never have to tell my agent, "Hey, go use this tool, open this MCP server." It goes and does its thing.
Getting to that feeling took a long detour through other people's memory tools, and one of them nearly cost me my whole setup.
The day a memory tool broke my whole setup
Building a memory layer was never the plan. I'm a marketer, and writing code isn't my thing. So my first move was to find one. The open-source community is amazing, and I assumed someone had solved this already. I went to GitHub and tried a couple of smaller tools, and they didn't really work. I tried building memory in Obsidian, and that didn't do it for me either. I tried the LLM wiki idea Andrej Karpathy talks about, a knowledge base the AI keeps in plain files, and it's good, but it wasn't the kind of memory I wanted.
Then I tried claude-mem, one of the biggest stars among AI memory tools, and it broke my entire Claude setup. I work only in my terminal, and the minute after I installed it, I opened a new Claude session and got a black screen of death. I tried again, and again nothing happened.
I spent the entire day on a hunt to save my setup. Thank God I do backups, so everything was back by the end of the day, but I was sweating, and I decided I wasn't taking my chances like that again. If the memory I wanted didn't exist in a form I could trust, I would build it myself.
Memory that maintains itself and costs almost nothing
Once I decided to build, the first question was what to build differently. Anyone who asks about my business plan raises the same worry: memory for AI is a huge red ocean, and every time you blink there are ten new players. It's true, and I'm fine with it. Since I started building, I've studied a lot of those tools to see how they think, and two problems kept showing up.
The first is maintenance. Almost all of them need you to keep the memory organized before your agent can remember anything, even the advanced ones with complex AI workflows built in. You have to work inside the right folder, keep everything up to date, and make sure the relevant information sits at the top. With time you become the maintainer of your own knowledge and memory, or of the setup you created to maintain it. So I made Total Recall automated and proactive: it captures and organizes on its own, and the agent reaches for it when it needs to.
The second is cost. Many memory systems lean on an AI model to sort and search your memories, and every one of those calls burns tokens, the units AI providers bill you for every bit of text a model reads or writes.
I'm not a developer. I don't work at a company where I can spend thousands of dollars in API calls.
So I built Total Recall to be super efficient. All the indexing and searching happens in plain code, with zero LLM calls. When your agent searches its memory, the only tokens you pay for are the agent's own question. I compared it with a memory built from markdown files, like those LLM wikis, and Total Recall saves up to 97% of tokens versus markdown-based systems. Tom picked up on why that matters: people are used to treating context as a layer of intelligence, and they're only now starting to treat it as a layer of memory that has to be efficient too. In the middle of the token maxing craze, when companies brag about burning through tokens, I'd bet that a year from now that idea will look ridiculous.
Under the hood, Total Recall runs a hybrid search. It combines semantic search, which finds memories by meaning, with lexical search, which finds the exact words, plus a few other signals. People sometimes call it a new kind of RAG. There are some RAG mechanics in there, but the big difference is that Total Recall knows how to forget. Each memory gets weighed by how recent and how important it is, so the important ones come up front, while anything that shouldn't pop into context gets almost archived. Fewer searches go nowhere as a result. And the agent gets progressive disclosure, seeing a little at a time and opening the details only when they're needed. I love progressive disclosure so much that if I ever had a rock band, that would be its name.
Every one of these choices came out of the same process, and most of that process is failing.
Try, fail, benchmark
Search is a layered problem with a lot of players, so people often want to know how someone outside engineering makes those architecture calls. My process starts with research and a close look at existing solutions, especially the ones that failed me, because by the time I sat down to design anything I already knew what to avoid.
I think probably the most important thing in building a product is understanding what not to do.
From there I test, fail and test again, and I fail a lot, really. I'll try a local model to see how it runs, measure it, and move on when it doesn't hold up. Tom, who comes from engineering, recognized the loop: something broke, I dug into why, chose a different path and kept going until I hit the target I had set. He said it's very close to how he would approach the problem himself.
The target comes from academic benchmarks, standard exams any memory system can sit, which let me compare Total Recall with what big players like Mem0 have published. On LongMemEval, the leading memory benchmark, Total Recall scores 98% end to end with an AI model as the judge, 490 of 500 questions, so I'm pretty much killing it right now. Some people dismiss benchmarks altogether, and I get the instinct, because trust in AI tooling is low. But the benchmark is my answer to that. I ran the test, I got the result, and that result is the only claim I make.
Benchmarks tell me the memory works today. Where memory goes next is a bigger question, and it's about who owns it.
Now we can
I think memory will soon be built into every agent and every harness you work with, and it will feel smooth. Whether the big companies let agents share that memory is less certain. Right now everyone is racing to beat everyone else, so we're in a very siloed stage, and your memory in one tool doesn't follow you into the next. Total Recall already works across them. It's the same memory for all your agents: talk to Codex, and Claude knows exactly what you discussed and why you made the decision you made. Eventually memory becomes basic infrastructure, and the models and agents run on top of it.
You also have to decide where your memory lives. The OpenAIs and Anthropics of the world can own that layer for you, or you can own it, and I think most of us would want to own it. So Total Recall is completely local and lives only on your machine. Maybe one day there will be a cloud version for people who prefer that, and even then the memory stays yours.
Tom keeps a list of 927 AI tools to review. He looked closely at about 120, and by his estimate around 15% of those were built by people who aren't trained engineers. He calls us people from the trenches, experts in our own field who end up building the tools we need, and who might come up with the idea no engineer thought of. I came from marketing, felt one problem every single day, and built the infrastructure to fix it, and I've done the same with Token Optimizer, an open source tool that has crossed 30K users in just a few months, and with Miss Chief, the little virtual assistant who helps me with a lot of my work. What made it possible was knowing the problem better than anyone, and knowing from the tools that failed me exactly what not to do. A few years ago, that wasn't enough for someone like me to build infrastructure.
Now we can.
Key moments
- 01:34"Big confession, I am not a developer": ten lines of Python, around 2018, then she gave up.
- 02:11Fifteen years in marketing, then "the hell with it," she quits her job to learn AI for real.
- 02:51Where Total Recall started: getting tired of re-explaining herself to agents.
- 05:57claude-mem breaks her Claude setup and costs her a full day.
- 11:32Zero LLM calls for search, and up to 97% fewer tokens than markdown files.
- 15:17The most important thing in building a product: knowing what not to do.
- 16:1798% end to end on LongMemEval, 490 of 500 questions.
Alex Greenshpun is an AI strategist and builder and co-founder of 10x Company. She builds Total Recall, Token Optimizer and Outsourcerer.
