Mesmo com GraphRAG, o agent se perde sem contrato de memória

The headline, "Even with GraphRAG, the agent gets lost without a memory contract," highlights the importance of a memory contract for an agent using GraphRAG (Graph Retrieval-Augmented Generation) technology. The source material provides four key points about this issue.

Firstly, the agent initially pulls a good-looking snippet, showing high similarity and confidence. However, it was using the wrong stub, indicating a poorly configured approach rather than weak RAG functionality. The high cost of the index and garbage in the input were also identified as problems. The previous post discussed querying the code graph, where the agent's performance drops when it asks the map instead of when the gitignore is correct.

Secondly, the market debate revolves around retrieval (lookup) versus synthesis (answering broader questions). Vector technology excels at the former, while GraphRAG is better at the latter, albeit more expensive to maintain. A hybrid approach can route queries effectively. The snapshot shows that ranking is not the deciding factor, but rather whether the agent can mix daily tasks with runbook content in the same context.

Embedding a "correct" answer in a dead guide can lead to policy issues. In this setup, the contract came before the vector, emphasizing that the vector alone may not be sufficient.

Thirdly, the proof lies in the agent's ability to cite the hub or just similarity scores when answering. The question is about the rule of entry into the agent's memory. A costly approach involves reconstructing memory by searching everywhere for "memory" in README.bak and other sources. An inexpensive method involves reading the hub first and then using grep if needed, with the response indicating which hub the answer came from. Without knowing the entry point, there is no memory contract.

Fourthly, the contract consists of four key elements, applicable to any vault. These include a single source, unique canonic pages, and scopes that define what each box can contain. Scopes include personal, operational, meta, and other categories. The rest of the content should be an echo or garbage. The memory should have least privilege, not relying on Pinecone features. Finally, the fifth and final point is a reminder not to include legacy, garbage, or identifiable company-specific content as inputs to the memory.