The source describes T-Search as an open agent-retriever built on Qwen3.6-35B-A3B and trained entirely on synthetic data. It returns ranked evidence chunks with rationales, not a completed answer. The release also includes TRuST, described as a manually assembled Russian benchmark for complex agentic search, plus SynthComp in Russian and English.
Its operating mechanism is a multi-round ReAct loop: the agent searches a corpus, explicitly saves useful chunks and coverage state, then begins the next round with selected memory instead of the full tool-call history. The source says each round has roughly a 32,000-token budget; once 75% is filled, search is blocked and the agent must either finalize or save evidence and advance. The harness also requires at least five searches and one saved chunk before an early round transition.
The mechanism is explicit evidence preservation across bounded contexts, combined with constraints that discourage empty resets and premature completion. The immediate relevance is that teams building retrieval systems are confronting context noise as a systems-design issue, not only a model-selection issue. The source is a vendor-authored technical account and does not independently establish comparative production performance, cost, or generalization beyond its reported setup. For product teams, the implication is that retrieval architecture may need separate evaluation from answer-generation quality. Watch whether the released model, code, and benchmarks enable independent replications that compare this memory strategy against summarization, full resets, and larger-context baselines.