Memory Soda: A Self-Hostable Memory Layer for AI Agents

Memory Soda is a POC that provides a memory layer to enhance AI agents' capabilities. It focuses on storing and retrieving information crucial to maintaining context and consistency in conversations over time.

• Intelligent Storage: Memory Soda processes conversations, identifying statements worth retaining, resolving them against existing knowledge, and preparing prompt-ready text for subsequent model calls.

• Multi-Layered Memory: Memory Soda utilises three layers of memory:

  • Working Memory: A live conversation window that is automatically compacted.

  • Episodic Memory: Summaries of conversation segments, with provenance for each fact.

  • Semantic Memory: Durable subject-predicate-object facts with associated time windows, enabling precise retrieval and handling of contradictions.

• Transparent Pipeline: A bundled playground visualises every request, extracted triple, and contradiction verdict, offering full transparency into the memory process.

Memory Soda is currently pre-1.0 and in early development, with known limitations documented for users.