general-agentic-memory

To build long-term memory for multi-hop reasoning, you use a dual-agent architecture where a MemoryAgent constructs just-in-time abstracts from documents for iterative on-demand retrieval. This enables structured memory and guided research at query time.

Just-in-time memory construction is a process where a MemoryAgent builds abstracts from documents and stores associated pages. It allows AI agents to retrieve context on demand rather than loading entire documents into the context window upfront.

An AI agent performs iterative retrieval for complex research by executing a cycle of planning, searching, integrating, and reflecting. A ResearchAgent orchestrates this to answer multi-hop questions over stored memories.

Yes, the dual-agent memory framework supports BM25 keyword retrieval, dense vector retrieval, and page-index retrieval. This flexible retrieval mechanism covers diverse information needs during deep research.

The best way to handle document-driven knowledge integration is using a duo-agent workflow that builds abstracts and performs iterative retrieval. This targets memory-centric workflows for long-context questions and document-driven reasoning.

You should avoid just-in-time memory optimization if your workflow does not involve multi-hop reasoning or long-context questions. It requires a dual-agent architecture, which may be unnecessary for simple, single-turn queries without stored memories.