Epistolary Quest

The simplest setup is what most people have actually talked to. A prediction engine — software that looks at the text in front of it and predicts the text most likely to come next — is given a system prompt describing the character, then the conversation so far, with the user's most recent letter at the end. The engine predicts the reply.

This is the original chatbot architecture. It works well for the first few exchanges. Try the demo: the early messages feel right. The character sounds like the character.

The trouble starts as the conversation grows, or as the user asks about anything specific. The chatbot cannot look anything up. Whatever the character is supposed to know has to be packed into the system prompt up front. Loading enough information to handle a wide range of possible questions runs into a hard limit: prompts have a maximum length. Past that limit, the engine truncates — oldest content drops to make room for new. The conversation gets summarized, compressed, or lost.

When the character does not have what they need to answer, their best option is to confidently guess. People often call this confident-but-wrong answer a hallucination, though the term gets used for a wider range of failures than any single definition captures.

The first attempt to extend the conversation past the truncation limit is to maintain a running summary alongside the recent messages. The system keeps a record of who the user is and what has been discussed. On each new turn, the prediction engine sees the summary plus the most recent letters. Older detail gets compressed into the summary or dropped.

To get the character knowledge the prompt cannot hold, the system can store it externally and retrieve relevant pieces on demand. The system has a library of documents about the character's world. On each turn, it searches the library for anything related to the user's question and pastes the matching passages into the prompt. This pattern is commonly called retrieval-augmented generation.

The next step is to give the library structure. For a character in a persistent world, the graph consists of who they know, what interactions they have had with each character, where they have been, what events they have participated in or know about, and what ideas they hold. Each of these is recorded as data, and the connections between them — who taught whom, who disagrees with whom, what event led to what — are recorded as edges between the entries.

The world is mapped out, persistent, and searchable by meaning rather than by exact-word match. When the user's letter comes in, the system runs a search by meaning over the graph and pulls in the relevant entries.

The first four architectures share a shape: the system pulls some context up front, hands it to the prediction engine, and gets one shot at a reply. This architecture changes that shape.

A prediction engine works well when the right information is in front of it. When it is not, the engine has two usual behaviors: refuse to answer or admit ignorance (both rare in practice), or give the most likely prediction it can produce without any way to verify it. The result is the same — a guess presented as an answer.

The change is to add an outlet: something productive the character can do other than give up or make something up. They can actually go find out. The character is given a description of the world they live in and a set of tools for exploring it. They can search the world for what the user has brought up. They can pull up the full record on a specific person, place, or event by name. They can trace the path between two things. They can recall what they have learned about the user from earlier letters. They choose which tools to use and in what order.

A reply is no longer one prediction. It is the result of a series of investigative steps: search, read, follow a connection, search again. Each step puts more of the right information in the engine's context, so each next step has more to go on. Investigation is the default. The character only writes when they have what they need.

The final piece needed to get a character that works reliably is to have something review the outputs and give the character feedback. After the character composes a draft, a separate scoring pass reads it and rates it across 13 specific dimensions: how well the voice matches the character, whether it answered what the user actually wrote, whether anything in it conflicts with what the character would know, whether it leaves the user room to disagree, whether the character has stayed in character or slipped into addressing the user as an AI assistant would, and others.

When a dimension falls short, the character does not just see "rejected." They see what fell short and why, and they treat that as feedback. They go back to their tools, look up what they need, revise the draft, and submit it again. The loop continues until the reply passes. The user only ever sees a passing reply.

The full picture: a structured world graph, search by meaning, the character investigating with their own tools across multiple steps, a scoring pass on every draft, and a revision loop the character drives when something falls short. The voice is still a prediction. The world, the connections, the continuity across letters, and the quality of every reply are grounded in a system that does not rely on prediction alone.