When Should an Agent Use Team Experience? TencentDB Agent Memory Adds JEV Decision Support

说明与用途

JEV 在技能回调测试中将令牌使用量降低约40%，保持任务成功率不变，并将困难任务的通过率从30%提升至40%。

问题

Agents may either ignore useful team assets or load irrelevant memories, wasting tokens and effort.

结果

JEV reduces token usage by ~40% in skill recall tests and improves pass rates on complex tasks, while keeping task success stable. Memory generation costs drop 13.8% and L3 calls fall 39.7% with high semantic synchronization (96.7%).

工作流

• 1. Agent receives task and candidate assets.
• 2. JEV receives candidate identifiers and descriptions and selects relevant assets.
• 3. System fetches and loads selected assets into context.
• 4. Agent executes task using loaded assets and records outcome.
• 5. After execution, JEV evaluates whether to regenerate long‑term memory (L3) based on changes in L2.
• 6. If JEV decides to update, system regenerates L3; otherwise it skips.
• 7. Execution records and feedback are fed back into Memory Hub for asset improvement.

证据

Repeated input-token injection after failed matches fell from 24.96 million to 15.04 million with JEV, a reduction of about 40%.

Task success was nearly unchanged: 66.7% with the main model making the selection versus 65.3% with JEV.

On a selected set of longer, more difficult SWE‑bench Pro tasks, the reported pass rate rose from 30% to 40%.

Total cost fell 13.8%. L3 generation calls fell 39.7%. Final semantic synchronization reached 96.7%.