Are these Agentic Memory Power Rankings correct?
Based on the comments, these are the best ways for your agents to remember things.
- Graph-based Systems (Graphify, Graphiti, TypeGraph)
Stores knowledge as connected nodes (entities/facts) and edges (relationships). Agents can query relationships, reason structurally, and incrementally add new information as "episodes".
- Markdown + Vector DB Hybrid (Obsidian + embeddings or similar)
Plain markdown files for readable notes + vector embeddings for fast semantic search by agents. Simple, local, and human + agent friendly.
- LLM Wiki (Karpathy-style + tools like usealmanac)
LLM automatically builds and maintains a structured, interlinked markdown "wiki" from your sources. The wiki itself becomes the persistent knowledge base instead of raw retrieval every time.
- Mem0
Dedicated memory layer that extracts, consolidates, and stores key facts/preferences from conversations. Provides personalized, long-term memory across sessions with easy add/search APIs.
- Agent-Native Memory Systems (e.g. CerebroCortex)
Purpose-built frameworks designed specifically for agents. Often use hierarchical or graph memory to let agents remember, learn, replay histories, and evolve over time.
- Hybrid RAG Approaches
Combines multiple methods (vector search + graphs + SQL/structured data) for more accurate and deterministic retrieval than pure vector RAG.
- Temporal / Time-based Memory
Tracks not just facts, but changes over time (what changed, when, and why). Adds provenance and evolution tracking on top of other systems (often graph-based).
I should say, many people are mixing approaches (e.g. graphs + vectors, or markdown + vector DB). I currently use Obsidian, but I'm curious to start using more graph-based systems for larger client projects.
Let me know in the comments your thoughts on the power rankings.