Feat: Introduce Graph-Based Context Indexing for Multi-Agent Shared Memory

[FenjuFu]

Motivation

With the rise of Graphify-Labs/graphify trending today, the community is clearly shifting towards structured context management for complex AI tasks. graphify demonstrates the value of converting codebases into queryable knowledge graphs to solve the "context window limitation" and "memory loss" problems in large-scale projects.

In multi-agent orchestration (our core strength in astron-agent), the current context passing mechanism relies heavily on linear text or flat vector search. As workflows become more complex (e.g., coordinating 3+ agents for code refactoring or financial analysis), agents often lose the "global view" of the project structure or business logic, leading to redundant work or hallucinated dependencies. We need a way to inject structured, graph-aware context into the agent's working memory, not just raw text chunks.

Proposed feature

Add a Graph-Based Context Indexing module to the Astron Agent runtime:

1. Graph Construction: Integrate a lightweight graph extraction step (similar to graphify's approach) that parses project files, API schemas, or domain documents into a node-edge graph (e.g., using Neo4j or a lightweight in-memory graph like networkx for smaller scopes).
2. Hybrid Retrieval: When an agent needs to understand a complex task, the system should:
   • First, query the graph for structural relationships (e.g., "Which modules depend on auth-service?").
   • Then, retrieve relevant text chunks based on the graph nodes.
   • Inject this structured context into the agent's prompt as a "Global Context Block" before task execution.
3. Multi-Agent Shared Knowledge: Allow multiple agents in a workflow to share this graph index, enabling them to "see" the same structural dependencies without re-parsing the codebase.

Why it fits this repo

astron-agent is positioned as an enterprise-grade SuperAgent platform. Enterprise workflows often involve large codebases or complex business logic where structural understanding is critical. By adopting the "structured context" paradigm highlighted by graphify, we can significantly improve the reliability of long-horizon multi-agent tasks, reducing errors caused by context fragmentation.

This feature complements existing Role-Based SuperAgent Orchestration by adding a cognitive layer to the orchestration, ensuring agents have a shared, structured understanding of the domain.

References

• Graphify-Labs/graphify: Demonstrates the power of code-to-graph conversion for structured AI context.
• iflytek/astron-agent: The target platform for this enhancement.

[prabhjeet16]

This resonates with a challenge we're seeing across AI-driven development workflows.
As codebases grow, the bottleneck often isn't code generation, it's maintaining a reliable understanding of architecture, dependencies, and business logic across the system. The issue you described highlights how context fragmentation can cause agents to lose the global view, leading to redundant work, dependency mistakes, and inconsistent reasoning across long-running tasks.
We've been exploring this problem at JigsawML. Rather than adding more generated context, we're focused on helping teams visualize structural relationships, dependencies, and change impact so developers and AI agents can understand how the system fits together before making decisions.
The idea of graph-based context indexing is interesting because it aims to solve the same underlying problem: preserving architectural understanding instead of relying solely on limited context windows.
I'd be curious: do you see the bigger challenge as context-window limitations, or as maintaining a shared understanding of the system architecture across agents? That's exactly the type of workflow we're researching, and I'd be happy to share a free demo and get your feedback.

[FenjuFu]

两者都重要，但对我们更难的是在多个 agent 之间维持共享的架构理解——上下文窗口限制只是让这种碎片化变得可见而已。图式上下文索引之所以吸引人，是因为即使单个 agent 的窗口装不下，结构关系本身仍然保留下来，可被反复查询。感谢你的思考。

[prabhjeet16]

That's an interesting perspective, and I agree that the architecture-understanding challenge exists independently of context-window size.

What stands out is that once multiple agents begin reasoning from different fragments of information, maintaining a shared model of dependencies and relationships becomes more important than simply expanding the amount of context available to each individual agent.

The idea of preserving structural relationships as a persistent layer of knowledge is especially compelling because it allows agents to repeatedly reference the same architectural truths instead of reconstructing them for every task.

We've been exploring similar challenges at JigsawML around dependency visibility and shared system understanding, so it's interesting to see the same theme emerge in multi-agent orchestration. It'll be interesting to see how graph-based context evolves alongside retrieval and memory strategies as these workflows become more complex.

Thanks for sharing your thoughts.