Halucinuje vám AI? Mně ne!

🔐Exclusive content on Patreon: Patreon: JirkaHernik ⚙️Best VPS for your agents: https://aox.cz/vps Yesterday you spent two hours working through a problem with AI, explained your whole project, and fixed its mistakes. Today you open it back up and it has no idea what you're talking about. And it won't even say "I don't know" - it calmly hands you an answer that sounds great and just isn't true. This is episode three of the AI pro lidi (AI for People) series, and it's all about why this happens and what you can actually do about it. First, let's get clear on what AI "memory" actually is. It's made of three things: the model's weights (what it learned during training), the context (what's in front of it right now) and external memory (files and notes it reloads next time). The model itself remembers nothing between requests - the app resends the context from scratch every single time. Then we look at why a bigger context window doesn't mean a better result. I run /context and walk through what's actually sitting in that window: the system prompt, memory files, skills, history and the auto-compact buffer. I show graphs from context rot research, the needle-in-a-haystack test, and the classic lost-in-the-middle U-curve, where primacy and recency hold up fine but the middle gets lost. Next I break down /compact: what it does, what its internal prompt looks like, why it's lossy compression (after a few rounds you end up with a summary of a summary), and when to reach for /new instead. And because important rules don't belong in chat history, we go through CLAUDE.md and AGENTS.md: what belongs in them (environment, architectural decisions, conventions, acceptance criteria, fail loudly instead of a silent fallback), what doesn't (a map of your folders, a README rewrite, class descriptions), and why every unnecessary word burns tokens in every single conversation. I also explain why I don't recommend /init: the generated file usually just duplicates what's already in the project, and as the code ages it starts actively confusing the agent. The study Evaluating AGENTS.md (Gloaguen et al., ETH Zurich and LogicStar.ai, 2026) found that context files generally didn't improve success rates and raised inference costs by more than 20%. The last third is about hallucinations. A language model isn't a database of truth, it's a generator of probable text continuations, and post-training evaluations can end up rewarding guessing: "I don't know" scores zero, a guess has a shot at a point. The counter to that is grounding - anchoring the answer in verified sources. At the end you get eight practical rules, including a ready-made escape-hatch instruction that gets the model to openly admit what it couldn't verify. SOURCES AND LINKS FROM THE VIDEO Claude Code - context window: https://code.claude.com/docs/en/conte... Claude Code - memory and CLAUDE.md: https://code.claude.com/docs/en/memory Claude Code - /context, /compact, /init commands: https://code.claude.com/docs/en/commands AGENTS.md (format for other tools): https://agents.md/ Chroma: Context Rot (technical report, 18 models): https://research.trychroma.com/contex... Lost in the Middle: How Language Models Use Long Contexts: https://arxiv.org/abs/2307.03172 Found in the Middle: Calibrating Positional Attention Bias: https://arxiv.org/abs/2406.16008 Evaluating AGENTS.md (ETH Zurich and LogicStar.ai, 2026): https://arxiv.org/abs/2602.11988 Why language models hallucinate (OpenAI, 2025): https://openai.com/index/why-language... Language Models (Mostly) Know What They Know (Anthropic): https://www.anthropic.com/research/la... Boris Cherny (creator of Claude Code) on Y Combinator: Boris Cherny: We Cut 80% of Claude Code’s ... PREVIOUS EPISODES OF THE AI PRO LIDI SERIES 01 - AI Agents from Zero: What You Need to Know: AI agenti od nuly: Tohle potřebujete vědět 02 - I Hired an AI Agent. Here's How You Can Set One Up Too: Zaměstnejte AI agenta. Ukážu vám jak! WHERE TO FIND ME Discord: Discord: discord Patreon: Patreon: JirkaHernik CHAPTERS 00:00 Intro 01:27 A genius with amnesia 04:25 What the /context command shows 06:19 Context rot 09:05 Lost in the middle 11:23 How /compact works 14:21 Compact always costs you something 15:46 Rules belong in a file 16:52 The exploratory tax 17:55 A tattoo from Memento 21:07 Why bans don't work 23:39 Why not to use /init 26:29 Try deleting CLAUDE.md 28:22 Why AI makes things up 30:16 Grounding: go check 31:47 Eight practical rules 34:06 Key takeaways #AI #ArtificialIntelligence #AIproLidi #hallucinations #ClaudeCode #context #CLAUDEmd #AGENTSmd #prompt #productivity This video has an English audio track spoken by an AI voice clone of the author. Audio created with Boson AI's Higgs Audio - https://www.boson.ai/higgs-audio