Trace an answer to the code¶
The graph is what makes an answer traceable. Every claim the Agentic Wiki writes lands on a symbol, a file and a note in one indexed structure, so a reader follows an answer back to the code and a stale claim can be found and retired.
Every indexed repository also ships a Graph view over that structure, with a folder collapse so a large repository stays readable. From any symbol you jump to the page that documents it. Agentic Search renders its own capability graph in the same view.
The view is an illustration. The graph under it carries question answering from a single file lookup to repository scale.
Three layers, one graph¶
The graph is a multiplex of three layers, and the console has a toggle per layer.
- Structural (AST) layer. Files, classes, functions, methods and interfaces are the nodes. Call, import and definition relationships are the edges.
- Entity layer. The subsystems, components, and concepts that matter in the codebase, extracted after the AST phase. Each is grounded in real AST symbols and source prose, never in generated text. Anchor edges connect an entity to the symbols it describes.
- Memory layer. Atomic notes anchored to both structural symbols and entity nodes.
Grounded by a code memory graph¶
Most documentation tools chop a repository into text and search it like prose. Mewbo builds a graph instead. The three layers are separate dimensions over one store, which is what a multiplex code memory graph means.
A question does not pull a handful of snippets that merely look similar. Keyword and semantic search seed the entry points, then the graph expands them across multiple hops, from a symbol to its callers to the module that owns them to the note explaining why.
Because every note is anchored, a stale answer cannot hide. When the code under a note changes, the next refresh checks that note and retires it if it no longer holds.
Why multiple hops matter
A flat search answers a question that spans modules with three disconnected results. Traversal follows real relationships instead. Semantic embeddings only accelerate this. Without them keyword retrieval still seeds the graph and the wiki keeps working.
The memory grows as the wiki is used¶
The memory layer is not frozen at indexing time.
- The indexer seeds it while writing pages, with anchored facts about each subsystem.
- Question answering feeds it. A durable fact from a run is condensed, anchored and merged off the critical path. → Question Answering
- You and your agents can teach it. Submit a fact the code alone will not reveal and Mewbo condenses it, drops duplicates, and merges it. Every later answer draws on the accumulated set.
Contribute an insight
Your agents take that last path as they work, through the MCP server's
submit_insight tool or the REST endpoint.