When I use an agent over a collection of documents, I have to decide where useful context should live and who maintains it. The model, the session and the harness do different jobs. These are approaches I have been experimenting with or checking against, not four interchangeable kinds of memory inside the model.
- Recurrent inference state. Mamba, state-space models and hybrids compress context into a bounded state during inference. That state belongs to the model’s execution. Whether a serving system retains or reuses it across requests is a separate runtime decision; it is not automatically a durable document store. Gu and Dao’s Mamba paper describes one such architecture.
- Reasoning text. The model generates a scratchpad while working through a question. The harness decides whether to keep that text in session history or supply it again later. Paying for those tokens is different from maintaining a persistent, source-backed corpus.
- Flat retrieval. Vector similarity and BM25 keep memory in an external index. I can update the source collection without retraining the model. Retrieval still depends on the index and the quality of the representation; storing a document does not guarantee finding the relevant part.
- Graph retrieval. Nodes, edges and communities add structure to external memory. The partitioning method matters because it determines which groups and relationships retrieval can follow. Microsoft's GraphRAG uses Leiden community detection; Nuthatch takes a stochastic block model (SBM) approach. Those choices are part of the method, not an interchangeable graph-shaped decoration. Leiden community detection nuthatch
There is also a commercial incentive worth questioning. More rounds of work mean more tokens sold. That does not establish why any particular laboratory chooses a memory design, but it is a reason not to assume that my preferred cost and persistence trade-offs are the provider's priorities.