Math Brain: Shared, Auditable Memory for Multi-Agent Mathematical Research
Year: 2026
Type: preprint
Abstract: Language-model agents now do a good deal of research mathematics, but what one session establishes is lost to the next: a counterexample found at noon is unknown to the agent that takes up the same problem at one o'clock, and corrected statements and failed routes vanish with the session that found them. We describe Math Brain, a shared memory built to keep them. It is a store of mathematical cards (definitions, theorems with their exact hypotheses, sources, and the agents' own proofs, counterexamples, computations and corrected statements) joined by typed links; on 22 October 2026 it held 13,306 cards and 19,362 links, 231 of them retracted and kept in the record. Agents reach it through a hook that serves a few cards on each prompt and stays silent on weak matches, a recall tool that prints whole cards with their identifiers and links, link strengths that grow with joint use, and a write-back gate that lands a packet of new cards only when a held-out recall test shows that it costs no answer. We give the structure and the ranking in full and prove the invariants on which safe sharing rests: append-only revision, order-independent bundle digests, acyclic supersession and dependency closure, and the separation of association from evidence. In paired tests with the same model, budget and questions, the memory supplied what the agents could not otherwise know. On 1,000 items whose answers are cards in the store, scores rose from 538 to 1,000, with no harm observed. On a second bank of 1,000 items run in three arms, the brain, the research repositories without the brain, and the model alone, Claude agents scored 95.4%, 82.0% and 18.9% on the 988 items run in all three arms; on the items that the first two arms both finished in time the observed rates were close (99.3% and 99.4%), so the brain's lead over the repositories is time, and it came at a quarter of the input and under three fifths of the session time and price. Codex agents, given more time, scored 95.2% with the brain and 97.6% with the repositories (p = 0.12). On 100 questions about results held in the store, agents scored 0 with no tools, 75 with text search and 83 with the recall tool, and adding the hook and the link walk gave 82 while cutting uncached input by a factor of 1.8. A ranking change raised held-out hits from 298 to 317 of 584 (exact McNemar p = 0.0019). On 61 reasoning and audit items the full package, which also adds a clock and a hint, beat tools alone 52 to 41 (p = 0.019, not significant after correction for the paper's nineteen comparisons). We report the nulls and losses with the same care: no significant gain over agents given the documents and a recall tool, none on thirteen new open problems, and a significant loss by an early version to agents given the documents, on 300 hard items. The question banks remain private; after release, repetitions on publicly answerable items are regression checks, not fresh held-out evidence. The paper also records mathematics from the same program, including a reduction of the Fourier entropy-influence conjecture: for every Boolean function H(f) ≤ 2(I(f) − Var f) + 2U(f), where U compares each Fourier weight with the weight above it, so the conjecture is equivalent to sup U/I < ∞. We prove this and a summed restriction criterion that would give the conjecture with an explicit constant.
Author: Shmalo, Yitzchak
Institution: Hebrew University of Jerusalem
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225911150
