Open Ontology of Polymer Chemistry: Accumulation and Self-Improving Memory Mechanisms for Polymer Research

Memory systems for AI agents are a valuable component of long-running workflows. They are also a strong form of lock-in for model providers, because the memory system that makes a model better with use may be inaccesible to users. An ideal memory system would accumulate both factual knowledge and the relationships of that knowledge to recurring tasks performed by the researcher. People accumulate particular and general knowledge through experience, the latter often through periodic reflection on what is working and isn’t working at a higher level of abstraction. Here, we introduce an open-source harness for polymer research meant to heuristically mimic (i) knowledge accumulation across research tasks, (ii) the review of new knowledge against existing knowledge, and (iii) performing periodic updates to task-knowledge relationships through reflection. The system, which we call the Open Ontology of Polymer Chemistry (OOPC, pronounced ‘oopsie’), is initialized with a knowledge base consisting of an open-source wikipedia with entries for most known polymers. From this starting point, the knowledge base was enriched by running a 1,092 task campaign across different areas of polymer chemistry, producing 8,899 distinct findings extracted from 1,679 sources on 157 different polymers. GPT-6.1 Sol and Opus 5.5 comparisons were run with and without the accumulated memory from this campaign on 60 held-out tasks with varying overlap in terms of polymer and properties being researched. The performance improvements tracked with research topic similarity to prior work researched from the harness, with token reductions of up to 19% and 48% and web search reductions of up to 80% and 72% (10.3 to 2.1 and 7.6 to 2.1 per answer) for Opus 5.5 and GPT-6.1 Sol, respectively. The harness was tested through five independent restarts of 500 tasks each with reproducible performance gains. An independent review of the performance caught rare contradictory claims and occasionally unsubstantiated claims depending on the underlying model (up to a few percent of all claims by claude on withheld tasks, none by codex). The reflection heuristic was tested as an update to the knowledge ontology after the accumulation period. These updates mainly consisted of extended lists of property names with marginal improvements to finding recall. The roles of review and ontology updates require further study as after knowledge accumulation these tests had little room to show appreciable gains given the consistently strong evidentiary basis for findings within the harness. Looking forward, domain-specific memory systems are likely to be valuable and ideally transferable research artifacts that can be shared at the group, community, and organizational levels.