SodaMemA self-evolving, agentic memory layer

Every memory can name the turn it came from, and knows when it stopped being true.

Retrieval that costs nothing

search and build_context make zero model calls. build_context returns a prompt-ready block with citations already assembled and budgeted.

Facts that expire instead of vanishing

Corrections are ADD-only. A superseded fact closes with a valid_until and stays readable, so "what did we believe last March" is still a question with an answer.

A foreign key, not a similarity score

FactEvent to SourceSpan to RawTurn. When someone asks where a stored fact came from, there is a row, not a ranking.

Where to start

you want to	go to
embed memory in a Python app	Getting started · Python
give Claude Code or Cursor memory	Coding assistants
run the HTTP service for a team	Self-hosting
wire up LangGraph, CrewAI, or another agent framework	Agent integrations
look up a route, tool or flag	Reference

What this is not

SodaMem is not a vector database with a memory-shaped wrapper. The parts that matter are the data model — a fact that points at its source turn and carries four independent time axes — and a retrieval path that is deterministic and free. Everything else in these docs follows from those two.