Memory-first
Concentric memory rings with retrieval filaments.
Verified against arxiv.org
A memory-first agent puts retrieval at the centre: before it reasons it pulls the relevant records from a store, and it writes structured memory back as it goes. The store, not the prompt, is the source of truth, so the agent's quality tracks the quality and freshness of what it has indexed. This is the shape behind research and knowledge agents, where grounding in a corpus matters more than raw reasoning. In anatomy terms the memory plane is load-bearing and the reasoning core orbits it.
The memory and retrieval plane is central; the reasoning core reads from and writes to it each turn; retrieval filaments (RAG, GraphRAG) feed the context; the perimeter governs what may be indexed and returned.
- ·Research and knowledge agents grounded in a corpus (OpenEvidence, AlphaSense, Elicit, Kensho).
- ·Enterprise question-answering over documents and databases.
- ·Long-running assistants that accumulate durable context.
- ·Grounded, citable answers tied to a source of truth.
- ·Improves as the index grows and is curated.
- ·Reduces hallucination versus prompt-only reasoning.
- ·Durable context across sessions.
- ·Only as good as the index: stale or thin data shows immediately.
- ·Retrieval quality (chunking, ranking) is a hard sub-problem.
- ·Write-back and deduplication add complexity.
- ·The retrieval hop adds latency.
Mapping classified by the Field Book from each agent's documented control flow.
Description, use cases and trade-offs: Field Book analysis, synthesised from established agent-architecture literature.
The corpus does not rewrite vendor documentation. Where copy is quoted verbatim it is marked as such and attributed to its source.