# SodaMem

A self-evolving, agentic memory layer for AI agents where facts get superseded rather than overwritten, entity profiles rebuild on demand, and recall never calls a language model, so the same question gets the same answer every time.

SodaMem is a memory layer for AI agents that treats facts as something that can be superseded rather than silently going stale: new information supersedes old, entity profiles rebuild on demand instead of drifting, and every answer traces back to the exact conversation turn it came from. Recall does not call a language model, so repeated queries are both free of that cost and deterministic.

It integrates with several runtimes: as a native, auto-injected plugin or an MCP tool for the DeepSeek Harness chat bot, as an MCP tool for other clients, as Python adapters for LangGraph, CrewAI and the OpenAI Agents SDK, as a TypeScript adapter for the Vercel AI SDK, and through a CLI and hooks for coding clients such as Claude Code and Cursor. The project reports scoring 92.8% (464 of 500) on the LongMemEval benchmark using a DeepSeek model as reader, planner and judge.

It is written in Python, distributed under the Apache-2.0 license, and can be self-hosted.

Website: https://sodamem.com