{"slug": "i-m-18-and-i-built-a-neuron-like-memory-system-for-ai", "title": "I'm 18 and I Built a Neuron-Like Memory System for AI", "summary": "An 18-year-old developer from Slovakia built a neuron-like memory system for AI that mimics human memory processes. The architecture includes five tiers, standby neuron agents that activate only when needed, and a neurogenesis mechanism that spawns new agents for emerging topics. The system passed all 324 tests across eight categories.", "body_md": "Not just RAG over chunks. Not vector search. Real memory — the kind humans have.\n\nThat question grabbed me and wouldn't let go. So I read about the hippocampus, Ebbinghaus forgetting curves, complementary learning systems, slow-wave sleep replay. And then I built it.\n\n## The Architecture\n\n┌──────────────────────────────────────────────────────────────┐\n\n│ 5-TIER MEMORY ARCHITECTURE │\n\n├──────────────────────────────────────────────────────────────┤\n\n│ │\n\n│ TIER 1+2 EPISODIC BUFFER Brain: Hippocampus │\n\n│ ═══════════════════════════════ Speed: <1ms │\n\n│ 64 working + 256 episodic items │\n\n│ Ebbinghaus decay: n^0.3 · e^(-λt) · importance │\n\n│ Forget threshold: 0.05 | Promote threshold: 0.65 │\n\n│ Access-based reinforcement on every read │\n\n│ ↓ │\n\n│ TIER 3 SEMANTIC STORE Brain: Neocortex │\n\n│ ═══════════════════════════════ Speed: ~50ms │\n\n│ ChromaDB v2 · all-mpnet-base-v2 (768-dim) │\n\n│ Hybrid search: dense + BM25 → Reciprocal Rank Fusion │\n\n│ ↓ │\n\n│ TIER 4 KNOWLEDGE GRAPH Brain: Association Cortex │\n\n│ ═══════════════════════════════ Speed: ~100ms │\n\n│ spaCy NER + 30 keyword patterns │\n\n│ NetworkX + SQLite · Multi-hop reasoning │\n\n│ Auto-relation inference: uses/works_on/depends_on │\n\n│ ↓ │\n\n│ TIER 5 COLD ARCHIVE Brain: Distributed Cortex │\n\n│ ═══════════════════════════════ Speed: async │\n\n│ Filesystem JSON · YYYY/MM organization │\n\n│ Full-text search · Thaw to active · Compact summaries │\n\n│ │\n\n├──────────────────────────────────────────────────────────────┤\n\n│ CONSOLIDATION PIPELINE (Sleep Analog) │\n\n│ ═══════════════════════════════════ │\n\n│ Decay → Cluster → Merge(LLM) → Rescore → Promote → │\n\n│ FindRelations(LLM) → Archive → Neurogenesis │\n\n│ Quick: 60ms (every 5min) | Full: ~3s (on idle) │\n\n├──────────────────────────────────────────────────────────────┤\n\n│ STANDBY NEURON AGENTS │\n\n│ ═══════════════════════════════════ │\n\n│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │\n\n│ │ Personal │ │ Tech │ │ Projects │ ...N agents │\n\n│ │ Agent │ │ Agent │ │ Agent │ │\n\n│ │ │ │ │ │ │ │\n\n│ │ DEEP 💤 │ │ LIGHT 🟡 │ │ DEEP 💤 │ │\n\n│ │ 0 RAM │ │ ~3KB RAM │ │ 0 RAM │ │\n\n│ │ 0 tokens │ │ ready │ │ 0 tokens │ │\n\n│ └──────────┘ └──────────┘ └──────────┘ │\n\n│ │\n\n│ Wake: trigger patterns + centroid similarity │\n\n│ Vote: all agents score → top K form consensus panel │\n\n│ Sleep: return to idle after task (zero token consumption) │\n\n│ Spawn: Neurogenesis creates agents from memory clusters │\n\n│ Prune: inactive agents auto-removed after 30 days │\n\n└──────────────────────────────────────────────────────────────┘\n\n## The Two Novel Things\n\n### 1. Standby Neuron Agents\n\nHere's what hit me: biological neurons don't all fire at once. Only ~2% are active at any moment. The rest wait, silent, consuming almost nothing.\n\nSo I built agents that work the same way:\n\nThey wake on trigger pattern matching + embedding similarity, vote in consensus panels (all agents score the query, top K get activated), communicate via sparse blackboard, and return to sleep.\n\n### 2. Neurogenesis\n\nWhen the system notices a cluster of memories forming around a new topic — say, 6+ memories about Minecraft — it *automatically spawns* a new specialized agent for that domain.\n\nIf an agent hasn't been woken in 30 days? It gets pruned.\n\n**Nobody else does either of these.**\n\n## Sleep as a Feature\n\nThe consolidation pipeline runs like biological sleep — reorganizing, merging duplicates, strengthening important memories, letting the rest decay:\n\nQuick mode: 60ms, zero LLM calls, runs every 5 minutes\n\nFull mode: ~3 seconds, LLM-powered merge + relation discovery,\n\ntriggers on idle detection (15+ minutes of inactivity)\n\n## Does It Work?\n\nEpisodic Buffer 41/41 ✅\n\nMemory Integration 32/32 ✅\n\nSemantic Store 26/26 ✅\n\nKnowledge Graph 37/37 ✅\n\nConsolidation 31/31 ✅\n\nStandby Agents 42/42 ✅\n\nCold Archive 27/27 ✅\n\nCross-Tier (E2E) 88/88 ✅\n\n────────────────────────────\n\nTOTAL 324/324 ✅\n\n## Why I'm Sharing This\n\nI'm 18, from Slovakia. This started as a random vibecoding project — a voice assistant. But the memory problem grabbed me and wouldn't let go.\n\nThe long-term thing that drives me: I believe better memory for AI could eventually help with conditions like Alzheimer's. Computational memory prosthesis. That's the direction I want to explore.\n\n**GitHub:** [github.com/FogyXT/JARVIS](https://github.com/FogyXT/JARVIS)\n\n**License:** AGPL-3.0\n\nThanks for reading.", "url": "https://wpnews.pro/news/i-m-18-and-i-built-a-neuron-like-memory-system-for-ai", "canonical_source": "https://dev.to/fogyxt/im-18-and-i-built-a-neuron-like-memory-system-for-ai-3hg6", "published_at": "2026-06-13 22:52:58+00:00", "updated_at": "2026-06-13 23:31:02.798777+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "large-language-models", "ai-agents", "ai-research"], "entities": ["Slovakia", "ChromaDB", "spaCy", "NetworkX", "SQLite", "Ebbinghaus", "Hippocampus", "Neocortex"], "alternates": {"html": "https://wpnews.pro/news/i-m-18-and-i-built-a-neuron-like-memory-system-for-ai", "markdown": "https://wpnews.pro/news/i-m-18-and-i-built-a-neuron-like-memory-system-for-ai.md", "text": "https://wpnews.pro/news/i-m-18-and-i-built-a-neuron-like-memory-system-for-ai.txt", "jsonld": "https://wpnews.pro/news/i-m-18-and-i-built-a-neuron-like-memory-system-for-ai.jsonld"}}