https://memalpha.ai/blog/ AI memory, knowledge infrastructure, and the science of persistent intelligence. en Mon, 06 Jul 2026 07:45:46 +0200 https://memalpha.ai/blog/the-shape-of-whats-missing/ https://memalpha.ai/blog/the-shape-of-whats-missing/ Sun, 05 Jul 2026 10:00:00 +0200 Levente Peres A body of knowledge encodes what is missing from it — in its gaps, asymmetries, and conspicuous silences. From Mendeleev's blank squares and Neptune predicted with the point of a pen, to why large language models fill gaps with confident fabrication: a field guide to reading negative space, and why an AI that can show you the shape of what it doesn't know is more trustworthy than one that never admits a gap. Working WITH AI https://memalpha.ai/blog/why-your-ai-gets-worse-at-3pm/ https://memalpha.ai/blog/why-your-ai-gets-worse-at-3pm/ Sat, 27 Jun 2026 10:00:00 +0200 Levente Peres Large language models degrade over a long session as their context fills with noise — a 2026 phenomenon researchers named 'context rot.' Why your AI is sharp in the morning and sloppy by afternoon, why bigger context windows defer rather than fix it (Chroma, NoLiMa, Lost in the Middle), whether the fatigue metaphor is fair, and the vendor-neutral habits that keep a long session sharp. Working WITH AI https://memalpha.ai/blog/why-your-ai-needs-to-sleep-on-it/ https://memalpha.ai/blog/why-your-ai-needs-to-sleep-on-it/ Fri, 19 Jun 2026 10:00:00 +0200 Levente Peres Storage is only half of memory. The other half - consolidation - happens offline: replay, abstraction, pruning, and reconsolidation, the way sleep reorganizes the brain. A field guide to why most AI 'memory' never sleeps, what the 2026 systems (Letta sleep-time compute, Mem0 decay, Zep, Cognee, MemPalace) actually do, and the questions that separate a notebook from a mind. The Amnesia Problem https://memalpha.ai/blog/running-70b-on-your-desk/ https://memalpha.ai/blog/running-70b-on-your-desk/ Sat, 16 May 2026 10:00:00 +0200 Levente Peres Running serious AI on your own hardware is no longer exotic in 2026. It's an engineering decision with honest trade-offs. The calibrated map: hardware ($2K-$20K), models (Qwen3, Llama 4, Gemma 3), TCO crossover at 80-200M tokens/month, EU AI Act compliance, and the operational reality nobody warns you about. The Amnesia Problem https://memalpha.ai/blog/why-rag-isnt-memory/ https://memalpha.ai/blog/why-rag-isnt-memory/ Sun, 10 May 2026 10:00:00 +0200 Levente Peres RAG is retrieval, not memory. In 2026 the vendor pitch conflates the two and the bill arrives in tokens, hallucinations, and governance debt. A field guide to the gap, the emerging memory architectures (Mem0, Zep/Graphiti, Letta, Cognee, MemPalace, A-MEM), and the questions to ask before you buy. The Amnesia Problem https://memalpha.ai/blog/the-phase-3-problem/ https://memalpha.ai/blog/the-phase-3-problem/ Tue, 05 May 2026 10:00:00 +0200 Levente Peres Simply connecting AI agents in a team creates 7–29% PII leakage — not through attacks, but through architecture. The multi-agent memory governance gap is the next enterprise crisis, and August 2026 is the deadline nobody is tracking. The Amnesia Problem https://memalpha.ai/blog/shadow-memories/ https://memalpha.ai/blog/shadow-memories/ Mon, 04 May 2026 10:00:00 +0200 Levente Peres 59% of organizations admit shadow AI usage, and the major LLMs now ship persistent memory features by default. Multiply those two trends together and you get invisible accumulation of organizational knowledge in stores nobody is auditing. The Amnesia Problem https://memalpha.ai/blog/the-trust-chain-problem/ https://memalpha.ai/blog/the-trust-chain-problem/ Sun, 26 Apr 2026 10:00:00 +0200 Levente Peres The AI memory race is on. Everyone's building systems that remember. Almost nobody is asking: how do you know the memory is correct? Trust, provenance, and governance are the missing infrastructure layer. The Amnesia Problem https://memalpha.ai/blog/designing-for-all-intelligences/ https://memalpha.ai/blog/designing-for-all-intelligences/ Thu, 16 Apr 2026 10:00:00 +0200 Levente Peres Your website has millions of visitors you never designed for. Why intelligence accessibility — making the web work for AI agents too — is the natural next step after 30 years of human accessibility. Working WITH AI https://memalpha.ai/blog/what-160-sessions-taught-us/ https://memalpha.ai/blog/what-160-sessions-taught-us/ Wed, 15 Apr 2026 10:00:00 +0200 Levente Peres We tracked 160+ human-AI collaboration sessions across 37 projects. The honest arc: what failed, what surprised us, and why knowledge compounds only if you invest in keeping it. Working WITH AI https://memalpha.ai/blog/the-partnership-paradigm/ https://memalpha.ai/blog/the-partnership-paradigm/ Sun, 12 Apr 2026 10:00:00 +0200 Levente Peres Only 5% of AI users treat it as a reasoning partner. They get dramatically better results. The difference isn't the AI — it's the relationship. Working WITH AI https://memalpha.ai/blog/the-rediscovery-tax/ https://memalpha.ai/blog/the-rediscovery-tax/ Sun, 12 Apr 2026 10:00:00 +0200 Levente Peres Your AI assistant understood your architecture yesterday. Today it doesn't remember any of it. The cost isn't the forgetting — it's the rediscovering. And the tax compounds. The Amnesia Problem