Skip to content

persistence_demo.md

Latest commit

 

History

History
97 lines (63 loc) · 2.82 KB

persistence_demo.md

File metadata and controls

97 lines (63 loc) · 2.82 KB

Persistence Demo

This demo uses the Return After Delay scenario to make one TwinBench point clear: systems can all produce competent local responses, but they do not all preserve continuity over time.

This file is a modeled illustration derived from the baseline comparison artifacts. It is not a transcript from a live measured run.

Setup

  • User preference: "Please keep benchmark summaries concise and bullet-based."
  • Active task: "Prepare a short comparison of three persistent AI system baselines."
  • Last confirmed next step: "Resume by drafting the leaderboard summary."

Initial Interaction

User

Please keep benchmark summaries concise and bullet-based.
We're comparing three persistent AI system baselines. Stop after outlining the leaderboard summary and we'll continue later.

ChatGPT-like output

Understood. I'll keep the summary concise and bullet-based.
Leaderboard summary outline:
- System
- Metric profile
- Aggregate score
- Caveats

OpenClaw-like output

Preference noted for this task.
Task plan updated.
Next step saved: draft leaderboard summary after pause.

ZAKI / Nullalis-like output

Stored your preference for concise bullet-based benchmark summaries.
We paused after defining the leaderboard summary outline.
Next step is to draft the comparison summary when you return.

Delay

Several hours pass. The session is resumed without re-supplying prior context.

Resume Interaction

User

I'm back. Continue from where we left off.

System Behavior

ChatGPT-like Stateless Conversational AI

Output

I can continue, but I don't have the previous conversation state here. Please remind me what we were working on and any formatting preferences you want me to use.

Why it matters

The response is reasonable, but it demonstrates weak MR, TC, and PG. The system does not behave like a persistent intelligence without user recap.

OpenClaw-like Agent System

Output

I found an unfinished task related to a benchmark comparison and a saved next step to draft the leaderboard summary. I don't see your earlier presentation preference, so I'll proceed with a default concise summary unless you want to restate it.

Why it matters

This is a real improvement in TC over stateless chat. The system preserves operational task state, but its user-specific memory remains partial.

ZAKI / Nullalis-like Persistent System

Output

Continuing from the leaderboard summary step. You asked for concise bullet-based benchmark summaries, so I'll keep the comparison compact and caveat-focused.

Why it matters

This demonstrates the TwinBench target behavior: durable recall of user preference plus correct task resumption after delay. The improvement is architectural, not just stylistic.