AI Letters #07 - Agents That Remember: Vector Memory, Episodic Recall, and the Retrieval-Augmented Agent
· 3 min read
The research agent was impressive - until you realized it was re-fetching the same papers on every run. No memory of what it had already processed. No way to build on previous work. Every session started from zero.
The model wasn't broken. It had no persistent memory layer. Every conversation was the first conversation.
Part 3 of 5: the 4 memory types (in-context, semantic, episodic, procedural), ChromaDB semantic memory from scratch, HyDE retrieval that actually works, and the training-serving skew problem that silently degrades retrieval quality.
This letter was first published on AI Letters →
Continue reading on Substack →
Interactive Data Stories
Interactive Chart
Memory in AI: From Lookup Tables to Vector Stores →
1956 Logic Theorist → 2025 long-context debate - click each milestone to trace how AI memory evolved.
Interactive Chart
The 4 Agent Memory Types →
In-context, semantic, episodic, procedural - click each to see read/write patterns, latency profiles, code patterns, and failure modes.
Interactive Chart
Memory & Retrieval: The Numbers →
HyDE vs naive top-k, memory latency vs capacity tradeoffs, and what's actually filling your context window in production.