Discussion

openJiuwen’s TTSE helps AI agents separate facts from learned tips

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#12507

openJiuwen has added TTSE, a framework that helps AI agents learn from earlier tasks without confusing facts about their environment with advice about what to do. Its reported benchmark results are promising, including a 97.8% success rate on ALFWorld with DeepSeek-V3.2, though they come from the team’s own paper rather than an independent evaluation.

Watch Desk analysis

What happened

TTSE, short for Two-Track Self-Evolution, separates declarative knowledge, such as where files are stored, from procedural tips, such as what action to take in a particular situation. The framework is now built into openJiuwen’s WorkSwarm office agent, according to a report by PANDaily citing QbitAI and the team’s paper.

The system starts with an empty rule base, learns from task traces and carries its rules into later tasks. It can retire misleading rules, keep a record of why they were retired, and turn contradictions into conditional tips. Anything it cannot confidently classify defaults to a tip rather than a fact. That is a sensible bit of housekeeping for agents that learn as they go.

Key findings

  • ALFWorld success reached 97.8%
    The team reports 97.8% with DeepSeek-V3.2, compared with 83.6% for the previous best method; with GLM-5, it reports 98.51%.
  • ScienceWorld improved
    Reported native success rose from 53.08% to 73.46%.
  • Two tracks beat either alone on GDPevo
    The full method scored 0.2616, against 0.0947 for zero-shot.
  • A PinchBench test showed a smaller gain
    Across 147 end-to-end tasks with OpenClaw, the reported average score rose from 0.7439 to 0.7822.
  • Office demonstrations used fewer tool calls
    One Excel data-cleaning task fell from 57 calls and 6 minutes 22 seconds to 20 calls and 1 minute 48 seconds after experience was added.

Why it matters

An agent that remembers a useful procedure can save time. An agent that remembers a bad assumption as fact can make the same mistake with admirable consistency. TTSE’s separation of environmental facts from conditional advice, along with its ability to retire and trace rules, addresses a practical problem in agents that carry experience between tasks.

The reported results span several benchmarks and office demonstrations, but they are not independent confirmation of performance in everyday use. The available report also does not say whether TTSE is offered as a standalone library. Users can try it through WorkSwarm 0.2.7 or later by enabling the TTSE setting.

Our read

The useful idea here is not simply that an agent remembers more. It is that the framework tries to keep different kinds of memory in their proper lanes, and to explain when it discards a rule. The benchmark gains make TTSE worth following, while the team’s own evaluation means real-world performance still has to do the talking.

What to watch

  • Whether TTSE becomes available outside WorkSwarm.
  • Independent evaluations of its benchmark and office-task results.
  • Whether rule retirement and traceability help prevent repeated mistakes in longer-running deployments.

Discussion spark: Should AI agents be allowed to carry learned tips between tasks by default, or should users have to review and approve what the system remembers?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.