Support teams for Sui applications have long suffered from a repetitive cycle of decoding identical transaction failures. Ticketwise, a project built for the Total Recall hackathon, attempts to break this loop by integrating on-chain memory with automated triage. The system reads failed transactions directly from Sui mainnet via Tatum’s gateway, maps abort codes to human-readable error names, and retrieves historical corrections stored in Walrus Memory.

Decoding the Chain

The core challenge in Sui support is that abort codes are opaque integers on-chain. Ticketwise addresses this by mapping 276 specific abort codes for major protocols like DeepBook v3, Bluefin Spot, and Cetus CLMM to their constant names in the source code. For example, abort code 5 in DeepBook's big_vector module is translated to ENotFound. When a ticket arrives, the agent checks Walrus Memory for a key like deepbook::big_vector#5. If a teammate previously taught the system the correct response, it uses that exact answer, complete with the instructor's name and date. If no memory exists, the agent explicitly states it doesn't know rather than hallucinating a fix.

Proof of Concept With Real Data

To validate the approach, the developer replayed 661 real failed transactions from DeepBook, Cetus, and Bluefin over a 13-day period ending November 10. In a controlled test, the first 435 transactions were used to teach one correction per distinct error. When triaging the remaining 226 transactions, the agent with memory disabled answered none of them with team-taught solutions. With memory enabled, it successfully answered all 226 using the correct meanings. This high success rate stems from the repetitive nature of support queues; the 13-day window contained only 10 distinct errors, all of which appeared in the initial training set.

On-Chain Memory vs. Traditional Databases

The project deliberately uses Walrus Memory instead of a standard SQL database to ensure team knowledge is not siloed on a single server. Every correction is stored as a blob on Walrus mainnet, verifiable via Walruscan links. The system includes a durability check that can wipe a search index and restore it from on-chain storage, a process that took 4.5 seconds in live testing. However, this decentralization comes with latency costs: writing a new memory takes approximately 20 seconds due to encryption and on-chain registration, while recall during triage takes about 3 seconds.

Key Takeaways

  • Ticketwise maps 276 Sui abort codes to human-readable names for DeepBook, Bluefin, and Cetus.
  • Walrus Memory enables shared, verifiable team knowledge that survives server resets.
  • The agent refuses to answer if it lacks a taught correction, preventing hallucinations.
  • Write latency is ~20s, but recall is fast enough for triage workflows at ~3s.

The Bottom Line

This is a pragmatic implementation of agentic memory that respects the reality of blockchain support: errors are repetitive, and trust requires verifiable history. By anchoring corrections on-chain, the team avoids the vendor lock-in of proprietary vector databases while maintaining a single source of truth for their support knowledge.