Tag
This tweet poses a technical interview question about deploying RAG chatbots at scale, explaining common pitfalls like state persistence and pointing to Akamai's GitHub repos and Developer Hub for reference implementations.
The author describes building a memory layer for news-reading AI agents to handle deduplication and improve continuity across sessions, using SQLite and vector storage with a Python API, CLI, and MCP server.
Argues that agent memory systems are evaluated on recall but not on whether recalled facts remain true; introduces a 'correctness under staleness' benchmark showing vector stores return stale facts 68% of the time.
L'article clarifie la différence entre le RAG classique (stateless) et la mémoire d'agent (stateful), et présente deux approches émergentes : Mem0 avec une mémoire à trois niveaux et auto-édition, et Letta avec une mémoire façon OS (core/archive). Il souligne que le vrai défi est la logique de mise à jour et d'oubli, pas le stockage vectoriel.
Argues that vector stores are unsuitable for storing business records (orders, balances, etc.) because semantic similarity does not guarantee factual correctness; recommends using SQL for structured data and vector stores only for unstructured information.
Oracle's AI Database now includes vector store functionality for embeddings-based image search, showcasing innovative features that make it a unified data storage solution.