Agent Memory Systems and Knowledge Graphs: Letta, Mem0, Graphiti, and Cognee
7 hours ago
- #knowledge graphs
- #benchmarks
- #agent memory
- Letta (MemGPT) uses self-editing text blocks without graphs, relying on the LLM to update memory via tool calls.
- Mem0 removed its optional graph layer in v3 and uses entity linking to boost retrieval without typed edges.
- Graphiti (Zep) focuses on bi-temporal fact graphs, preserving history by invalidating edges rather than deleting them.
- Cognee employs an extract-cognify-load pipeline, modeling memory as a graph with explicit stages for graph creation.
- Benchmarks like LOCOMO, LongMemEval, and HotPotQA evaluate recall and reasoning, but scores vary due to different eval setups.
- Graph-based systems incur higher ingestion costs but support traversal and historical queries, while simpler systems prioritize efficiency.
- The choice between graph and non-graph memory depends on the need for fact traversal and handling of updates over time.