Searching data using semantic similarity alone is flawed. Graph-aware retrieval is the way to go.
The basic implementation runs flat retrieval to select records that match a query. This ranks records by semantic similarity and returns the closest matches from the store.
This works well when you store the information you need in independent records, but it fails when you need to understand relationships between facts.
Graph-aware retrieval is a huge upgrade.
It has three components:
1. It stores facts as persistent memory records
2. It connects records with directed links labeled with their relationships. For example, "supersedes," "supports," or "contradicts"
3. It expands search results by going through the links and returning every connected record
Because of this, your query results will contain both matching facts and information about their relationships.
Imagine you have two stored conversations:
- "The meeting is on Monday."
- "The meeting has moved to Tuesday."
If you use a simple flat search to answer, "When is the meeting?", you'll get both records back because they match the question.
If you use graph-aware retrieval, you will get back a "superseeds" relationship showing that Tuesday replaces Monday.
Oracle AI Agent Memory 26.8 implements graph-aware search. You can read their technical walkthrough and check out the code here:
Thanks to the @OracleDevs team for partnering with me on this post.