Which do yall use? The Agent Traverse or Query. My take is traversal.
Since querying is fuzzy and noisy. Traverse get your agent the info they want and need, but not traverse aimlessly.
My approach is saving the why and what:
Why = The Edges (causal relationship between nodes)
What = The Node (events, decisions, issues)
Together it make the whole causal chains.
Where do the Agent start in the graph? Based on its current task/etc, the Agent can traverse and understand the topic fully or view the causal chains which contain the Root node to the leaf node.
The Graph evolve with your Agent and your Agent Experience Compound .
Repo (https://github.com/NodeDex/NodeDex-v0.1) run Locally.
Feel free to give a try and tell me what you guys think.
Solo-build 0v0
Top comments (4)
I think both have their place. Querying is usually faster and scales better, but as you said, it can become noisy and miss important context. Traversal is interesting because it preserves the relationships between events, decisions, and outcomes instead of retrieving isolated memories.
Personally, I like a hybrid approach: use querying to find the relevant entry points, then traverse the graph to understand the causal chain. That seems more reliable than relying on either one alone.
Niceee, what graph Db did you use, or you made one yourself?🧐
Interesting one. I don't think it's either/or — the hard part of traversal is picking where to start, and that's basically retrieval too. So I usually query to find a start node, then walk from there. How do you handle a question that touches two weakly connected parts of the graph?
Thanks for the correction , you're right its either, what i do is also query for the start, but the agent don't start at the node level, they start at the root level, think of it like using a file explorer a file is a root, and the name tell you everything you see what you get + a short description so the agent query and it returns the relevant roots + a couple entry nodes (matched on concept tags + the label path, rolled up to the root). For the two weakly connected parts , since the system saves what the agent experience, everything related to each other should have links , so if a chain of node hit two different root the link will bring the agent to the other root, or the agent can just query the whole chain. Does that answer your question?