Oracle × Channel Talk × Bloom: Agent Memory and Betting on AI

1/ Security matters more in the AI era. One breach and it's over. Context memory is the key, and today Taewan Kim, a director at Oracle, explained the idea really well.

2/ Our special speaker was Kyunghoon Lee, Chief AI Officer at Channel Talk (a Korean customer messaging and support platform). He shared the whole messy process of their transition and how they adapted to AI, holding nothing back. You can tell they've struggled and gone through a lot of trial and error: try this, and if it doesn't work, try that. It's still evolving, but there's a reason he chose to join a company he had invested in back when he was a VC.

3/ The best part of Claude Bloom is still the small-group sessions. Ironically, I don't join them, so I have no idea what people talk about, but people keep signing up anyway. I'm curious, so at the end I ask each group to present, and we always run 30 minutes over. The Claude Bloom community is full of people who love to talk.

4/ Oracle made limited-edition stickers for us, the catering was good, and the operations and space were clean and comfortable. Thank you to Sujin Hong for all the help. (On the Claude Bloom YouTube channel, in one of our videos on sneaking AI into your work, I showed that my emails are 100% automated, and she seemed a little let down. But it's true, so I have nothing to say. Still, the AI always carries what I really mean.)

5/ After the event, the three of us from Cliwant took the elevator down, and it felt like we were a hired crew leaving a kids' birthday party. It was fun, and we cracked each other up.

6/ On the way out, an attendee asked what Cliwant actually does, and I said we build an AI service for writing bid proposals. They said they'd always wondered. It's a long story. You row while the tide is in.

Oracle and Channel Talk have supported Claude Bloom since the early days. Thank you. The groups also posted their insights on the Claude Bloom Discord.

1. Why agent memory becomes a security hole

Oracle is a company that sells databases. Yet the keynote topic was AI agent memory.

The link is the files that pile up on your local machine. When you use Claude Code or run multiple agents, memory files build up locally to keep the context of your work, and there's more in them than you'd think. API keys and passwords go in, and if you've connected Gmail or company systems, internal information ends up there too. Data like this, unmanaged and at risk of exposure, is called shadow data.

For personal use it doesn't matter, but the moment you use AI for company work, it's a different story. Important company assets end up scattered across local files on someone's PC.

Oracle's answer was to manage the memory agents use inside Oracle Database. You install a single Oracle Agent Memory package and it plugs straight into existing OpenAI or LangChain frameworks. Because memory is stored in the database instead of local files, security gets stronger, and summarization and embedding save tokens. Even when multiple agents write records at the same time, transactions keep everything in order.

The governance side was impressive too. Companies that have adopted AI solutions now have an FDE (forward deployed engineer) come in once a month to review the past month's prompt usage and build the skills the team needs. That only works because every record stays in the database.

As AI gets more advanced, more people have no idea where information is being left behind. That gap is where the risk is.

2. Why he bet his career on AI

Kyunghoon Lee worked as a VC for seven or eight years and joined Channel Talk this January. He went to work at a company he had invested in.

While he was a VC, the trends kept changing. When blockchain took off, everyone invested only in blockchain startups, and then it was the metaverse. So when AI came along, he first figured it would fizzle out in two or three years too. That view changed in early 2025, when the CEO of Allganize, one of his portfolio companies, said, "AI won't replace people. People who use AI well will replace people who don't." He had never thought of it that way, but somehow he became convinced it was true, and from then on he stayed up all night vibe coding. He said it's a bigger shift than mobile, and he feels lucky that ten years into his career he's in a position to judge it and learn it.

Choosing the customer support market came from analyzing it as an investor. What LLMs do best is problems with clear answers and language-based work, and after coding, support was the field that fit best. Anyone can do it without a license, there are manuals, and it all happens through conversation. The Korean SaaS market is at most 2 trillion won, but the call center market is 10 trillion won.

The first thing he did at Channel Talk was automate the sales team's meeting notes. Reps had four meetings a day, and each one took 40 minutes, from writing up notes to entering them in Salesforce. Last July he built a workflow that records with Fireflies, processes with n8n, and enters everything into Salesforce automatically.

Getting to unlimited tokens took several steps. Creating a dedicated AX (AI transformation) team failed, because work piled up on them and they didn't know the day-to-day work. The next approach was to develop one person in each team who was good with AI. It's like running: you see a friend running and think you could do it too, so you start. That's how the culture slowly began to change.

The turning point was performance reviews. They added "What impact did you make with AI last quarter?" to the evaluation, and at the same time gave everyone in the company a personal Claude Code account. The next day, the team chat blew up with posts like "I built this" and "I built that."

All those requests led to unlimited tokens. People said even premium accounts didn't give them enough tokens. The reasoning was that an hour of someone who's good with AI is worth more than the tokens, so letting them use more beats making them wait. They opened it up to unlimited. Part of why the decision was quick: the CFO seat was empty at the time. Ha.

Once they opened it, non-developers came to 1 million won a month and developers to 50 million won a month. He asked the CTO whether one developer was really doing that much work, and the CTO said they were handling ten people's worth of work alone. These were mostly ICs in their mid to late twenties. A setup had emerged where one person with no leadership title runs the work of a ten-person team.

The case that stuck with me most was the barista. At the office cafe, orders started out spoken, then people wrote their names on paper. Then one day a tablet showed up in the cafe. It showed the menu, you added items to a cart and ordered, a receipt printed out, and a Slack alert went off. The person who built it was the barista, alone, with Claude Code.

His routine is one LinkedIn post a day, which takes an hour or two from research to writing. After joining the startup he went three weeks without writing and felt himself getting dumber. Handling only what was right in front of him left no time to think more broadly, and when you don't think, your ability to think shrinks. So he started writing again this week.

He said that if you hand in work that obviously reads like AI wrote it, people actually think less of you. In the AI era, the ability to boil things down to the core matters more. A short piece with only the essentials is worth more than a long document that smells like AI, and whether you can do that is what separates people.

Asked what humans should focus on in the AI era, he answered briefly: "We're Homo sapiens. We should focus on thinking." LLMs look like they can do anything, but once you build agents yourself, that illusion breaks fast. "Shaping them is really hard." The fact that terms like context engineering and prompt engineering exist at all is proof that LLMs don't just handle things on their own. Setting guardrails, feeding in context, and setting direction is still the human's job.

"People who are good with AI are, in the end, people who are good at their jobs." He meant people who can give clear instructions and clearly review what comes back.

"AI only learns from data that's already finished, but humans are always one step ahead, because we have to make new decisions today." That one sentence summed up the whole fireside chat.

Asked about AX at large companies, he said that when the CEO uses AI well, the organization uses it well. Companies where the CEO doesn't use it and just says "We have someone good at AI, go try it" don't do well. The people running the company have to use it themselves.

He disagreed with Dario's prediction of the end of white-collar work. Since AI arrived there's actually more you can do. Large companies may choose to use AI for efficiency, but Channel Talk chose to grow. They're hiring more right now, ran subway ads at Yeoksam Station, and are offering 30% higher salaries. When one person who's good with AI joins, things don't shrink. The amount of work you can put out grows.

The future Channel Talk sees is replacing people's work, not selling tools. They already have customers where AI handles 80% of support, and replacing just 10% of the 10 trillion won call center market is 1 trillion won. As his picture of an ideal future, he pointed to a company called Lesser, where the manager is an AI. You load the CEO's context into the AI, and the AI evaluates people and assigns work. The AI does the actual work, and people sit above it.

3. Nine groups, one conclusion

After the small-group discussions, group leaders presented one by one. On the first question, "How far can you trust what an AI agent produces?", nine groups from different industries and career stages reached a strikingly similar conclusion: AI executes, humans make the final call.

Only the level of involvement differed. Someone from finance said one error does so much damage that the human has to be heavily involved. Someone doing AI development at LG CNS (LG's IT services arm) said they hand almost 100% of the coding stage to AI. Someone working on embedded software said just getting the AI to understand the full context is hard. Still, the conclusion was the same.

The test was how much responsibility a decision carries. A team lead decides within a team lead's authority, a CEO within a CEO's, and a human has to make those calls directly. Put the other way, any step that carries no weight of responsibility can be handed to AI.

Sangwoo Moon, who works as an AX builder at a foreign consulting firm, said he's experimenting with delegating even QA review to AI. The AI reports what it got wrong, and he designs the next harness so it doesn't repeat that mistake. "If AI can make mistakes, so can humans. I'm building a workflow where the AI learns by correcting those mistakes."

4. The workflows people built

The second question was "How are you actually using AI right now?" Every group brought at least one example.

Automating delivery of in-house gift vouchers, recording every customer conversation and having AI generate a customer needs blueprint automatically, turning unstructured manufacturing process data into a single template in bulk, automating the expense process: everyone was already building something in their own field.

One that stuck with me applied the dropout idea from model training to context handling. It was an experiment in deliberately removing parts of the context at random to improve performance. Another group used several models for cross-validation, with one model reviewing another's output.

Someone doing AI work in construction said they built an app that makes 3D models from photos, demoed it to a client, and got a good response. They aren't a developer.

The most impressive moment of the day was when one person recorded their group's discussion on the spot, summarized it with an LLM, and used that for their presentation. They demoed their own workflow live at the event.

5. What to teach the next generation

The third question drew the most discussion. People with kids, people with juniors on their teams, and people who are new themselves each spoke from where they stand.

When Sunwoo Kim, a first-year college student majoring in data science, started presenting, the room broke into applause, and the logic was sharp. "Education today is about finding the set answer, and AI does that better. We need fundamental thinking skills. Unlike a calculator or GPS, AI makes us delegate thinking itself. The next generation actually needs to keep AI at a distance on purpose and build the ability to think."

Another group said the same thing in different words. Knowing how to write good Python or good JavaScript doesn't mean much anymore. How to speak, how to instruct, and how to delegate are now the core skills. Natural language education comes before coding education.

One person who trains new hires at their company said they hand a new hire a task and ask them to leave the URL of their conversation with the AI. They go through the conversation log together: "If you'd asked this way here, you'd have gotten a deeper result." It teaches the process of talking with AI, not the AI's output.

Jinbae Park, who came from a medical startup, said he made a game with his daughter. She drew sketches, he put them into Claude to digitize them, and she decided "this is right, this isn't," and the game came together. "Age doesn't matter. It's an era where if you can do it, you can try anything." That line stayed with me.

6. The people AI works well for

Several groups said something similar: as AI gets more advanced, the uniquely human roles, defining the problem and carrying final responsibility, only get heavier.

Yeonkyung Bae from Bloom Energy shared a story pointing the same way. When a team member doesn't just bring the AI's output, and you can see "they put this thought into it, they considered this part," that's when you can actually understand it together. People who are good with AI turned out to be people with clear thoughts of their own.

At the end, one person said that even in the AI era, relationships between people are what remain. The greeting you exchange with eye contact in the morning, those everyday interactions, will matter most, and people who are good with AI will be the same people who are good with other people.

Originally written in Korean on Brunch, May 30, 2026.