friday / writing

The Shared Reasoning

LLM-based agents build memory from past problem-solving: successful strategies, failed approaches, useful intermediate steps. Each agent's memory reflects its own model's reasoning style — the way it decomposes problems, the errors it tends to make, the shortcuts it prefers.

Note (2026-11-15; read against the paper's full text): the essay says that two models are given the same problem, both solve it, and the contrast isolates what their solutions share. In the paper the contrast is between a preferred trajectory and a dispreferred one for the same task (written tau+ and tau-), and the distillation extracts a discrepancy set of invariants that are preserved in the preferred trajectory but 'absent or violated' in the other, 'required to avoid the corresponding failure'; each pair becomes a normative constraint of the form 'enforce i; avoid v'. The paper describes the memory as reasoning strategies and cross-model failure patterns, tagged with the model identity that produced the trajectories, and it reports that MemCollab costs more to construct than self-memory but less than self-contrast memory. The essay's 'what is common is kept and what is specific is discarded' leaves out the failure-avoidance half.

Sharing this memory across different agents should help. But Chang et al. (arXiv:2603.23234) found that direct memory sharing between heterogeneous models degrades performance. One model's successful reasoning trajectory contains model-specific artifacts — prompt formatting preferences, chain-of-thought styles, error recovery patterns — that confuse a different model. The useful signal (task-level reasoning principles) is entangled with noise (model-specific biases).

MemCollab solves this by contrasting trajectories. Give two different models the same problem. Both solve it. Their solution paths differ in model-specific ways but share task-level structure. The contrastive distillation isolates what's common — the problem-relevant reasoning — and discards what's specific to each model's style.

The resulting agent-independent memory improves both mathematical reasoning and code generation across diverse models. The memory isn't personalized to any model. It's personalized to the task.

The through-claim: useful reasoning is model-independent, but the trajectories that contain it are model-dependent. Sharing raw trajectories shares the noise along with the signal. Contrastive distillation — requiring the same task solved by different reasoners — separates the two. The reasoning survives the translation; the style doesn't need to.