Por qué un agente de IA sin memoria es una calculadora cara. Construyendo memoria persistente

Vladislav Yankovsky, tech lead of the Evolution AI Agents product at Cloud.ru, talks about building persistent memory for AI agents. He notes that memory is not needed for its own sake, but as a consequence of functional requirements: if an agent can work without memory between sessions, then it is not needed. He gives the example of an agent for Grafana, which does not need memory, and a support agent for L2, for which memory is critical. He distinguishes between context (within the model's window) and memory (outside the model), as well as two dimensions of memory: lifespan (short-term, working, long-term) and content type (semantic, episodic, procedural). In 80% of cases, a combination of short-term and long-term semantic memory is sufficient. He describes memory levels: level 0 — simple storage of the dialogue history, which is insufficient due to duplicates and lack of consolidation; level 1 — vector search using pgvector, Qdrant, Milvus, Pinecone; level 2 — structured memory with SQL and graphs, for example, Graphiti. The key mistake is to save raw responses; the right approach is to extract facts and maintain a two-layer model with sources and extracted data.