Tools · LlamaIndex
LlamaIndex Memory for Agents
LlamaIndex provides composable memory for agents — chat memory buffers, vector memory and memory blocks — letting LlamaIndex users assemble persistence without leaving the framework.
LlamaIndex memory
Components
LlamaIndex memory components
| Component | What it does | When to use |
|---|---|---|
| ChatMemoryBuffer | Stores recent chat messages | Short-term session memory |
| VectorMemory | Semantic search over past interactions | LTM-lite within LlamaIndex |
| Composable memory blocks | Mix buffer + vector in agent workflows | Custom agent pipelines |
| Vector store integration | Plug any supported vector DB | Production LTM backing |
Comparison
LlamaIndex vs dedicated memory frameworks
| Approach | Strength | LOCOMO J | Best for |
|---|---|---|---|
| Engram | Managed Weaviate pipeline | N/A publicly | Weaviate shops |
| LlamaIndex DIY | Composable, in-framework | N/A (DIY) | LlamaIndex agent workflows |
| Mem0 | Managed extraction API | 66.9 | Framework-agnostic POC |
| LangMem | LangGraph native | 58.1 | LangChain shops |
FAQ
Frequently asked questions
What LlamaIndex memory types exist?
ChatMemoryBuffer (recent messages), VectorMemory (semantic search over history), and composable memory blocks in agent workflows. Integrates with any LlamaIndex-supported vector store.
LlamaIndex memory vs LangChain?
Both offer buffer + vector patterns. LangChain adds LangMem for LangGraph-native LTM. LlamaIndex composes memory in agent workflows. Engram/Mem0 work with either via HTTP.
LlamaIndex vs Mem0 for agent memory?
LlamaIndex = DIY composition in-framework. Engram = Weaviate-native managed pipeline. Mem0 = framework-agnostic managed extraction + retrieval API.
Is LlamaIndex enough for production LTM?
VectorMemory + external vector DB works for many workloads. For managed extraction, eval-backed pipelines compare Engram or Mem0 alongside LlamaIndex agents.
What vector store should LlamaIndex memory use?
Any LlamaIndex-supported store — Weaviate (Engram), Pinecone, pgvector, Qdrant. See vector databases.
LlamaIndex agent memory in 2026?
Composable buffers + vector memory for in-framework persistence. Pair with Engram/Mem0 for managed cross-session LTM. See add memory guide.