v0.6.4 Announcement

English

Agent Memory Bridge v0.6.4 is out.

This release expands reviewed calibration into slice-aware reporting before any wider assist-mode rollout.

What changed:

1. The reviewed calibration set is larger

• 16 labeled samples
• more real coordination, retrieval, runtime, memory-shaping, storage, and model-routing cases

1. Calibration is now slice-aware

• per-slice summaries
• clearer error shape instead of one blended score
• easier to see where classifier help is already stable and where it still drifts

1. Assist gating stays in place

• minimum_confidence = 0.6
• low-confidence classifier output stays visible in calibration and shadow mode
• but still does not flow straight into assist-mode enrichment

Current calibration snapshot:

• from the deterministic reviewed-sample harness
• classifier_exact_match_rate = 0.875
• fallback_exact_match_rate = 0.062
• classifier_better_count = 13
• fallback_better_count = 2
• classifier_filtered_low_confidence_count = 2
• retrieval is currently the loosest slice at classifier_exact_match_rate = 0.6
• 78 passed

This is the point of the release:

• stop treating calibration as one global number
• expose which slices are ready and which still need reviewed samples

GitHub:
https://github.com/zzhang82/Agent-Memory-Bridge

Release:
https://github.com/zzhang82/Agent-Memory-Bridge/releases/tag/v0.6.4

简体中文

Agent Memory Bridge v0.6.4 发布了。

这次发布把 reviewed calibration 扩成了 slice-aware reporting，目的是在继续放大 assist-mode 之前，更明确地知道“哪一类已经稳，哪一类还没稳”。

主要变化：

1. reviewed calibration set 更大了

• 现在有 16 个 labeled samples
• 包含更多真实的 coordination、retrieval、runtime、memory-shaping、storage、model-routing case

1. calibration 现在是 slice-aware 的

• 有 per-slice summary
• 不再只看一个混在一起的总分
• 更容易看出 classifier 在哪里已经稳定、在哪里还在漂

1. assist gating 继续保留

• minimum_confidence = 0.6
• 低 confidence 的 classifier 输出仍然会出现在 calibration / shadow mode 里
• 但不会直接混进 assist-mode enrichment

当前 calibration snapshot：

• classifier_exact_match_rate = 0.875
• fallback_exact_match_rate = 0.062
• classifier_better_count = 13
• fallback_better_count = 2
• classifier_filtered_low_confidence_count = 2
• 当前最松的 slice 是 retrieval，classifier_exact_match_rate = 0.6
• 78 passed

这版真正想做的是：

• 不再把 calibration 当成一个全局大数字
• 而是把“哪些 slice 已经 ready、哪些还需要更多 reviewed samples”直接暴露出来

GitHub:
https://github.com/zzhang82/Agent-Memory-Bridge

Release:
https://github.com/zzhang82/Agent-Memory-Bridge/releases/tag/v0.6.4