openJiuwen debuts TTSE, a two-track experience framework for agents

openJiuwen has released TTSE, an experience-learning framework for LLM agents that targets a familiar weakness: agents that finish tasks without getting better at them. Chinese tech outlet QbitAI reported the release on Nov. 10.

TTSE splits experience into two tracks. FACT records properties and regularities of the environment itself, favoring claims that can be verified directly and that hold across tasks. TIP records execution methods under specific conditions, expressed in the form “under these conditions, perform this action.”

Each track keeps its own numbering, presentation format and evidence trail, so a single piece of experience can be cited, checked and updated — and when a run fails, the team can trace which factual judgment or which piece of advice was responsible. Items that are hard to classify default to TIP, reducing the risk of mistaking unverified tactics for environment facts.

The framework starts from an empty rule base: the agent runs a first batch of tasks, and FACT and TIP entries are induced from those trajectories. Later batches load the previous round's experience, execute, and then run four stages in sequence.

The stages are Blame, Retire, Contradict/Synthesize and Induce. Blame mines failed trajectories for rules that may have misled execution. Retire moves those rules to a retirement pool while keeping the reason, batch and timestamp on record. Contradict/Synthesize looks for the hidden conditions behind apparently conflicting rules and merges them into a more precise TIP. Induce compares successful and failed trajectories to add, revise, confirm or remove entries.

The team evaluated TTSE on five benchmarks, repeating every experiment three times. On ALFWorld it reached a 98.51% success rate with GLM-5 and 97.8% with DeepSeek-V3.2, 14.2 percentage points above the previous best method, GraSP, at 83.6%, while using fewer steps.

On ScienceWorld, MACRO Reward rose from 0.7204 to 0.8534 and native success from 53.08% to 73.46%. On GDPevo the full two-track setup averaged 0.2616, about 2.8x the 0.0947 zero-shot baseline. On SOPBench, combined with Bayesian skill management, overall success climbed from a 79.15% baseline to 91.59% for a single-track variant and 94.58% for the fused two-track version.

For real-world tasks the team attached TTSE non-invasively to OpenClaw and ran 147 end-to-end tasks three times each: average score improved from 0.7439 to 0.7822, a 3.83 point gain that held in the same direction across all three runs, with the spread narrowing from 3.33 to 1.61 points.

The ablations are the interesting part — recording only facts or only tactics both underperformed the full two-track design. The method and results are written up in the paper “A Two-Track Online Self-Evolution Framework for LLM Agents.” The open question is whether the blame-retire-synthesize-induce loop stays stable on longer task chains and across more models.