Tier agent memory into working / episodic / semantic with compression for ~90% storage reduction

Indium·medium signal

A three-tier memory design keeps a small working memory (4–8K tokens), compresses finished interactions into discrete episodic memories at ~90% storage reduction, and distills recurring patterns into semantic memory. Retrieval routes by recency — temporal indexing for recent episodes, vector search for older ones — which beats a single flat vector store on temporal (+29.6) and multi-hop (+23.1) queries.