Zep provides long-term memory for AI using temporal knowledge graphs.
| Feature | Rice | Zep |
|---|---|---|
| Philosophy | Cognitive architecture mimicking human brain regions | Knowledge graph with structured data retrieval |
| Working Memory | Dynamic decay and attention scoring | Session buffer with summarization/windowing |
| Procedural Memory | Server-side WASM execution | None (relies on client-side tools) |
| Data Structure | HDC with graph equivalents | Temporal knowledge graph with nodes and edges |
| Retrieval | Separate APIs for past (reminisce) vs current (drift) | Hybrid semantic + graph + keyword search |
Rice uses a four-component cognitive architecture (Working, Episodic, Procedural, Semantic). Zep uses a temporal knowledge graph that explicitly models nodes and edges with bi-temporal tracking.
Rice implements dynamic decay and attention scoring. Items naturally fade if not accessed, and drift() returns items sorted by relevance. Zep uses session buffers with summarization and windowing.
Rice includes Procedural Memory for server-side WASM skill execution. Zep relies on client-side tool execution via agent frameworks like LangChain.
Rice separates "remembering the past" (reminisce) from "orienting current context" (drift). Zep combines semantic vector search with graph traversal and keyword search.