Every couple of weeks, there’s a new “memory layer” for AI agents. And honestly… most of them do the same thing. They take your conversation history, extract entities and relationships, and compress everything into a knowledge graph. Sounds neat. But there’s a problem. That’s lossy compression. You’re deciding what matters at ingestion time - before the agent even knows what it’ll need later. Anything that doesn’t fit the schema? Gone. Nuance? Flattened. Context? Partially lost. We hit this exact wall while building Vektori. So we took a different approach. Instead of compressing everything into a graph, we keep memory in three layers: L0 — Facts: high-signal, filtered, fast to search L1 — Episodes: patterns discovered across conversations, no rigid schema L2 — Raw sentences: untouched, never loaded unless needed That last layer changes everything. Nothing gets thrown away. If a voice agent needs to recall exactly what you said — tone, phrasing, context from a past session — it can reconstruct it. The graph doesn’t replace reality. It sits on top of it. Early results? 73% on LongMemEval-S. If you’re building voice agents, this matters more than you think. Because voice isn’t just data - it’s nuance. integrations related to pipecat would love to know how well it works for voice agents 😄 https://github.com/vektori-ai/vektori