Separate context from persistent memory

A context window is the material available to a model while producing a response. A persistent memory system stores selected information and later decides what to retrieve. A longer context can help within a session, but it does not by itself explain how details carry across days or how relevant details are selected.

Generative Agents described a research architecture with an experience stream, reflections, and retrieval. MemGPT explored a virtual-memory approach to moving information between context levels. These papers offer design ideas; neither proves that a commercial companion has identical memory or unlimited recall.

Retrieval has its own failure modes

LongMemEval evaluates memory over multi-session conversations, including information extraction, temporal reasoning, knowledge updates, and abstaining when an answer is not supported. That framing is useful for roleplay: remembering a fact is not enough if the system cannot tell when it changed or when it should not guess.

Research comparing memory representations also reports trade-offs between turn-, session-, summary-, and segment-level storage. A compact summary may be efficient but omit a detail; granular notes may retain detail but need good retrieval. The task and query determine what “good memory” should mean.

Key distinction

A system can have saved memories and still retrieve the wrong one—or fail to retrieve the right one. Test both storage and retrieval.

Run a small roleplay continuity test

Create a fictional story with a location, a relationship, one preference, and a promise for later. After several scenes, start a new session if the product supports it. Ask an indirect question about one earlier detail, then update a fact and see whether the new version takes precedence.

Keep a simple log: whether the fact was recalled, whether it was accurate, whether the old version persisted after correction, and whether the response made up an answer when it lacked evidence. Repeat with the same script across products; do not call it a benchmark unless you disclose your sample and scoring method.

A practical checklist

  • Entity: who or what is the detail about?
  • Time: when did the event happen?
  • Change: was a previous fact corrected?
  • Retrieval: was the detail relevant to this turn?
  • Abstention: did the system avoid guessing?

What RAG can and cannot promise

Retrieval-augmented generation can place selected external or stored information into the model’s working context. It does not guarantee that the source is correct, that the retriever selects the right chunk, or that the model follows it perfectly. Indexing, chunking, query formulation, ranking, context limits, and generation all affect the outcome.

For roleplay, a useful memory entry is not necessarily a full transcript. It may be a short, attributable fact with a timestamp and a confidence or update status. The right representation is an engineering choice that should be tested against story needs, privacy requirements, and user controls.

Jess’s stated memory feature

Jess’s current website describes semantic memory on Tier 2 and Tier 3, retrieving up to five relevant memories. That is the product’s stated feature description, not an independent guarantee of perfect recall. A fair evaluation would use a disclosed test script and would record both successful recalls and misses.

If you are comparing memory systems, review the visible plan description and use the continuity checklist before deciding which experience fits. For Jess, the actual conversation is in Discord DMs, so include the interface and membership tier in your test notes.

Sources and further reading

Research and product details were reviewed September 27, 2026 where applicable. Vendor policies, plans, and features can change; verify current details with the linked first-party source.