Built this as a side project to explore a problem I keep running into with agents: they lose useful continuity across threads and resumed work.

Pallium is a local-first memory sidecar. It stores selected evidence, derives compact memory like decisions/findings/checkpoints, and returns evidence-backed memory blocks for follow-up questions.

It is intentionally not a general knowledge base or vector DB. The focus is scoped memory for agent-mediated conversations.

Repo: https://github.com/rore/Pallium