I've found that a simple markdown knowledgebase (with some useful extensions like semantic search & git context) is all I need to improve the memory of my agents. Even my non-coding agents have a memory repo.

Here's my implementation: https://hraness.com/kb