Knowledge and memory
How Eldric reasons
A question to Eldric passes through several parts, and only some of them are language models. Eldric's memory, and most decisions about what to look at and what to allow, are not made by a language model. The words in and the words out are, and so is the choice of the next step in a chat or an agent task. What you see as reasoning is the two working together.
Done without a language model
- Remembering. Recall of facts, past tasks and learned methods compares your question with what is stored. Where an embedding service is available, a small embedding model turns text into numbers; otherwise words are matched. Nothing is generated.
- Putting memory together. Before an agent plans, the past tasks and methods found are ranked and cut to one short block by fixed rules: how close each one is, how recent, and what a person said about it.
- Learning methods. Repeated tasks become candidate methods by counting: enough checked episodes, enough successes, and one sequence of tools behind most of them. An administrator decides whether a candidate is used.
- Your verdict. "Did it work?" is recorded as you answered and weighed by a formula, never by a model's opinion of the reply.
- Forgetting. Compressing and deleting old episodes follows dates and settings.
- Which tools a chat is offered. Eldric compares the question with an index of what each tool is for and offers a short list instead of every tool. A model that cannot use tools is refused by rule.
- Experiments. Whether an A/B result is significant is decided by a chi-square test or Welch's t-test.
- Forecasts. A forecast comes from a forecasting model trained on time series, not from a language model. It reads your numbers and returns numbers with a range.
Done by a language model
- The answer you read, and any "thinking" text shown with it.
- Choosing the next step in a chat or an agent task: which of the offered tools to call, with what, and when to stop. Rules decide what is offered and what is allowed; the model picks from that.
- Text work: translating, summarising and the other text operations of Content transform.
- Eldric's own review of finished conversations, which writes down themes worth remembering.
Why it is built this way
Memory that a model writes can be wrong in ways nobody checks. So what Eldric keeps, how it ranks it and what counts as success are decided by rules you can read and settings you can change. A model's text is treated as evidence, not as the record. A task counts as successful when its tools say so or when you say so, never because the reply says it worked.