AI Memory Is Broken And How I Fixed It
summary
Built a voice AI agent after MeZero dropped names and GBT4‑mini hallucinated family details, then shipped a three-layer memory that actually works: a deterministic facts table that supersedes old facts (no LLM writes storage), a pinned profile blob cached in-browser for instant responses, and per‑exchange 3072‑d embeddings stored in pgvector/Postgres for fast similarity search. TypeScript validates candidate facts, CodeX speeds development, Recall Me / speaktome is open source, and the system stops calling a partner by the wrong name.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 3:42Building my AI memory app
- 6:13Memory framework problems
- 8:21Three-tier memory system
- 12:32Vector search for recall
- 13:56Fact validation with TypeScript
- 14:59Prototyping lessons
- 17:49Open-source study tool
- 21:40Handling conversation interruptions
- 23:23Q&A
related talks
23mYour Agent Has Amnesia: Why Mobile AI Agents Need Memory to Be Useful
01h 49mBuilding Memory Augmented Agents with MemoRizz
54mThe Anatomy of Memory in Humans and AI Agents
25mWhat does it take to build long-term state for AI agents?
18mArchitecting Agent Memory: Principles, Patterns, and Best Practices
46mThe Anatomy of Memory in Humans and AI Agents
58mAdvanced Autonomous AI Agents: Memory, Collaboration and Evolution in Action
15mMemory-Augmented AI Agents: Enhancing Long-Term Reasoning and Context Awareness