Observations
Research Notes
Short observations from ongoing research into long-lived LLM systems: questions, tensions and working principles rather than architectures or implementations.
-
Convergent Evidence
Before building, we review three kinds of sources in order: scientific literature, existing systems and working implementations, and community discussion. Where they disagree, the disagreement is recorded rather than smoothed over. Where independent sources agree, the conclusion is treated as the strongest one — provided the sources are genuinely independent and not retelling one another. Every decision is written down together with its reasons.
-
An Honest “I Cannot Know”
Evaluating memory means measuring what a model does without it, too. A system that answers correctly with memory but invents answers without it is not reliable; it is lucky. In our evaluations, tasks that depend on memory are also run without it, and the expected result is an explicit statement that the answer cannot be known — not a plausible guess.
-
Train on Proven Rules Only
Adapting a model to a set of rules is only as sound as the rules themselves. If rules are taught before they are verified, their mistakes become the model’s habits and are much harder to find later. Our order is deliberate: define the rules, prove the properties they must have, and only then adapt a model to work within them — while the memory continues to check every proposal.
-
Memory That Outlives the Model
Models will be replaced; a long-lived system’s history should not be replaced with them. If memory, commitments and rules are stored in a form tied to one model, every upgrade becomes a loss of continuity. We treat the model as a replaceable reader and proposer, and the memory as the part of the system that must stay stable, inspectable and independent of any provider.
-
The Model Proposes, the Memory Verifies
A model’s output is a proposal, not a record. In a long-lived system, text written by a model should not become a fact, a decision or an obligation just because it was written. A separate set of checked rules decides what is accepted, and the model’s contribution stays visible as the model’s own words. This keeps the history trustworthy even when the model is wrong, manipulated or replaced.
-
Revision Without Rewriting
Changing a conclusion in response to stronger evidence does not necessarily make a system inconsistent. For a long-lived LLM, a deeper challenge is preserving that an earlier position was genuinely part of its history while allowing a later position to replace it without rewriting the path between them.
-
More Reasoning Is Not Always Better Reasoning
Additional computation can improve some decisions, but it can also add repetition, noise or unnecessary complexity. An important question is not only how much an LLM can reason, but when further reasoning is actually valuable.
-
Continuity Beyond a Single Interaction
Longer context alone does not necessarily create continuity. An LLM may retain information while still failing to preserve coherent priorities, interpretations or unfinished lines of reasoning across time. Continuity may therefore require more than simply making additional history available.