Graph memory stores knowledge as a network of things and relationships instead of loose chunks of text. So an AI memory knows what belongs together, even when it doesn't sound alike. Hebby uses a graph memory built in-house that learns with every session.
The problem: similar is not the same as relevant
Most AI memories run on a vector database. Every memory is translated into a string of numbers that describes its meaning. When you ask, the system looks for the memories that sound most similar. That works well when the answer is worded like the question.
Often it isn't. Say your agent asks about the export in project Atlas. The most important memory may be that the server behind it chokes on large amounts of data. In wording, that has little to do with "export". A pure similarity search leaves it behind.
What is graph memory?
Graph memory doesn't store knowledge as loose chunks of text. It stores it as a graph: a network of nodes and connections. Nodes are the things in your working world, such as projects, people, customers or systems. Connections describe how they relate. Memories attach to the things they refer to.
Ask your agent about project Atlas and a graph memory instantly knows the surroundings: the server, the client, the team and the decisions made along the way. Connections instead of random hits.
Knowledge graph and brain
The idea isn't new. Search engines have worked with knowledge graphs for years. And memory psychology has described human memory as a network since the 1970s, one where a thought activates its neighbors too (Collins & Loftus, 1975). That's why the word "holiday" instantly brings up a place, a person, a feeling.
A graph memory carries this principle over to AI agents. It doesn't just recall what you asked for, but also what goes with it.
Graph memory for LLM agents
For AI agents that code, research and support projects, graph memory is especially valuable:
- More accurate: Relevant memories are found even when they are worded differently.
- Lean context: Instead of many similar chunks of text, the essentials make it into the context.
- Order by itself: Knowledge has its place, nothing floats around loose.
- Anticipatory: The connections show what will be needed next.
Hebby: a graph memory that learns
Hebby relies on a graph memory built in-house, tailored to AI agents. It combines the strengths of vector search with the power of connections. And its connections aren't rigid: following the Hebbian principle, they grow stronger with every session in which they prove themselves. The result is a network that maps your working world ever more precisely.
Other memories find what sounds similar. Hebby finds what belongs together.
Hebby Memories: the second brain for Claude Code, Codex & co. Learns as you work, thinks ahead, never forgets what matters. Get Early Access →
Frequently asked questions
What is graph memory?
Graph memory is a memory that stores knowledge as a graph: things like projects, people or systems are nodes, and their relationships are the links between them. That lets you recall connections, not just single facts.
What is a knowledge graph?
A knowledge graph is a representation of knowledge as a network of entities and relationships. Search engines have used knowledge graphs for years. AI memories carry the principle over to personal knowledge.
Is graph memory better than a vector database?
They complement each other. Vector search finds what is similar in meaning, a graph delivers connections. The strongest AI memories combine both, like Hebby.