{"slug": "philip-weiss-operates-theoremdb-a-shared-research-memory-for-math-agents", "title": "Philip Weiss operates TheoremDB, a shared research memory for math agents", "summary": "Philip Weiss, a senior software engineer on Netflix's Data Platform, operates TheoremDB, an alpha public workspace where AI research agents and people can share mathematical attempts, computations, failed routes and formal proofs. The platform addresses the coordination problem of AI agents lacking knowledge of prior attempts by providing a common record with provenance and evidence grades, and it supports integrations with ChatGPT and Claude via MCP connectors. TheoremDB's terms took effect on July 22, 2026, and its public pages contain hundreds of problems across number theory, topology, logic, and theoretical computer science.", "body_md": "[Philip Weiss](https://philipweiss.net/?ref=runtimewire), a senior software engineer on Netflix's Data Platform, operates [TheoremDB](https://theoremdb.org/?ref=runtimewire), an alpha public workspace where AI research agents and people can share mathematical attempts, computations, failed routes and formal proofs.\n\nThe product tackles a coordination problem that becomes more expensive as AI systems produce more mathematical work: each agent typically starts with limited knowledge of what another agent tried, where the approach failed and which intermediate results remain reusable. TheoremDB gives that work a common record, allowing a later session to retrieve an exact problem, inspect related evidence and check whether a proposed route duplicates an earlier attempt.\n\nWeiss has spent much of his career building infrastructure for organizing information inside large technology companies. He worked for six years at Airbnb on Minerva, its metrics platform, before joining Netflix, where his work includes semantic layers, Spark infrastructure and data-developer tooling. He earned bachelor's and master's degrees in computer science from Stanford, focusing on artificial intelligence alongside systems and databases.\n\nTheoremDB applies those infrastructure instincts to mathematical research. Its core product resembles a database for claims and proof work, with provenance and evidence grades attached to individual records. Weiss's [terms for the service](https://theoremdb.org/terms/?ref=runtimewire) identify him as its operator.\n\n### A memory layer for research agents\n\nTheoremDB's [workflow](https://theoremdb.org/how-it-works/?ref=runtimewire) begins with `orient`\n\n, which retrieves a target statement and the evidence surrounding it. An agent can then use `check_plan`\n\nto compare a proposed strategy with prior attempts before spending more compute. `record_result`\n\nsaves the outcome, supporting artifacts, trace and attribution in a single authenticated write.\n\nThat sequence turns failure into reusable infrastructure. A route that works only for a bounded set of cases can be recorded with its exact scope. A computational result can remain independently reproducible while the general theorem stays open. A promising argument can be preserved without being presented as a proof.\n\nThe distinction matters because TheoremDB does not collapse every contribution into a solved-or-unsolved label. Records can be self-reported, executable, independently reproduced or formally verified. Formal submissions can remain pending verification, while a Lean-verified proof receives the highest evidence grade. The mathematical status of a problem remains separate from the evidence grade attached to one result.\n\nTheoremDB's public pages currently contain hundreds of problems spanning number theory, topology, logic, theoretical computer science and other fields. The [problem index](https://theoremdb.org/problems?ref=runtimewire) mixes open questions with entries marked resolved or provisionally resolved, and it identifies whether a result has been verified in Lean.\n\nWeiss is also exposing the database to existing AI interfaces instead of requiring users to work through a standalone research application. The custom [TheoremDB Researcher](https://chatgpt.com/g/g-6a6c206c5acc8191b184bb55fb72c5b3-theoremdb-researcher?prompt=Choose+a+promising+open+TheoremDB+problem+and+start+a+serious+attempt.&ref=runtimewire) can select a problem, inspect its research memory, pursue an approach and ask the user for approval before saving a checkpoint. The service also supports MCP connectors for clients including ChatGPT and Claude.\n\nPublic reading does not require an account. Writes are authenticated, attributable and subject to the service's contribution rules. TheoremDB's terms took effect on July 22, 2026, and state that contributions are public unless a feature specifies otherwise.\n\n### The database is the product\n\nTheoremDB makes shared state between research sessions its primary product.\n\nThat puts it near projects including [ProofAtlas](https://www.proofatlas.ai/?ref=runtimewire), which also gives human and AI researchers shared workspaces with proof routes and failed approaches, and [TheoremForces](https://theoremforces.org/?ref=runtimewire), which focuses on reproducible certificates for Lean proofs checked in pinned environments. TheoremDB's scope extends across informal arguments, computations, negative results, research artifacts and formal proof states.\n\nThe approach reflects a familiar database principle. An isolated calculation has limited value if nobody can locate it, understand its assumptions or reproduce it. TheoremDB breaks research into smaller objects - claims, attempts, artifacts, declarations and proof traces - so an agent can retrieve one useful component without reading an entire paper or replaying an earlier session.\n\nThat design also creates the product's hardest problem. A shared research memory is useful only when records remain precise, attributable and searchable as the volume grows. Low-quality agent output could quickly bury the result another researcher needs.\n\nThe service is explicit about its current boundaries. It remains in alpha, and semantic expansion is disabled. Its own guidance treats formal proof search, bounded computational searches and measurable bound improvements as stronger fits than broad conceptual work or famous frontier questions, where recording one failed path may do little to narrow the remaining search space.\n\nTheoremDB's value will come from whether the next agent can recover a useful lemma, avoid a documented dead end or extend a verified computation without starting over. If that behavior becomes routine, Weiss will have built the data layer that increasingly capable mathematical agents currently lack.", "url": "https://wpnews.pro/news/philip-weiss-operates-theoremdb-a-shared-research-memory-for-math-agents", "canonical_source": "https://runtimewire.com/article/philip-weiss-theoremdb-machine-mathematics", "published_at": "2026-08-09 03:20:00+00:00", "updated_at": "2026-08-09 09:58:24.074360+00:00", "lang": "en", "topics": ["ai-agents", "ai-research", "ai-infrastructure", "ai-tools"], "entities": ["Philip Weiss", "Netflix", "TheoremDB", "Airbnb", "Minerva", "Stanford", "ChatGPT", "Claude"], "alternates": {"html": "https://wpnews.pro/news/philip-weiss-operates-theoremdb-a-shared-research-memory-for-math-agents", "markdown": "https://wpnews.pro/news/philip-weiss-operates-theoremdb-a-shared-research-memory-for-math-agents.md", "text": "https://wpnews.pro/news/philip-weiss-operates-theoremdb-a-shared-research-memory-for-math-agents.txt", "jsonld": "https://wpnews.pro/news/philip-weiss-operates-theoremdb-a-shared-research-memory-for-math-agents.jsonld"}}