Major - P3 TL;DR
The LlamaIndex Suite spans two open-source frameworks: LlamaIndex, for building RAG and agent applications, and llama-agents, its Workflows runtime and deployment server for running agentic workflows in production.
MongoDB Atlas is already supported across nearly every LlamaIndex storage area — vector search, document store, index store, key-value store, plus Voyage AI embeddings and reranking. This work closes the two remaining gaps: agent memory in LlamaIndex (its newer Memory feature currently works only with SQL databases) and workflow persistence in llama-agents (which today supports only SQLite, Postgres, DBOS, or LlamaCloud). We are adding a MongoDB option for both.
Future State
| Domain | What it stores | MongoDB — Before | MongoDB — After |
|---|---|---|---|
| Operations & State | LlamaIndex: document/node store, index metadata, ingestion de-dup cache. llama-agents: workflow handler status, event journal, tick log, workflow state | LlamaIndex: ✅ docstore, index store, and key-value store all run on MongoDB. llama-agents: ❌ SQLite / Postgres / DBOS / LlamaCloud only | LlamaIndex: ✅ unchanged. llama-agents: ✅ MongoDB now backs the workflow store |
| Memory | LlamaIndex: conversation sessions plus long-term memory (auto-extracted facts, semantic recall). llama-agents: — (no memory store of its own) | LlamaIndex: ⚠️ MongoDB backs only the deprecated interface; modern Memory needs a separate SQL DB. llama-agents: ➖ n/a — delegated to LlamaIndex Memory | LlamaIndex: ✅ MongoDB backs the modern Memory API directly (sessions, facts, semantic recall) on MongoDB alone. llama-agents: ➖ unchanged |