v0.6.3 Announcement

English

Agent Memory Bridge v0.6.3 is out.

This release deepens classifier calibration before any wider assist-mode rollout.

What changed:

1. The reviewed calibration set is larger and more useful

• 10 labeled samples
• classifier-vs-fallback winners per sample
• missing-tag and extra-tag analysis
• average match score, not just exact-match rate

1. Assist mode now has a confidence gate

• minimum_confidence = 0.6 in the public config
• low-confidence classifier output can stay visible in calibration and shadow mode
• but it no longer has to flow straight into assist-mode enrichment

1. The benchmark and signal foundations stay intact

• memory_expected_top1_accuracy = 1.0
• file_scan_expected_top1_accuracy = 0.636
• signal lifecycle still holds claim -> extend -> ack / expire / reclaim

Current calibration snapshot:

• classifier_exact_match_rate = 0.9
• fallback_exact_match_rate = 0.0
• classifier_better_count = 9
• fallback_better_count = 1
• classifier_filtered_low_confidence_count = 1
• 78 passed

This is the point of the release:

• make assist rollout answerable to reviewed evidence
• keep low-confidence model output from silently changing durable memory

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

Release:
https://github.com/zzhang82/Agent-Memory-Bridge/releases/tag/v0.6.3

简体中文

Agent Memory Bridge v0.6.3 发布了。

这次发布把 classifier calibration 再往前推了一层，目标是让后续 assist-mode rollout 更可控，而不是只靠感觉。

主要变化：

1. reviewed calibration set 更大，也更有判断力

• 现在有 10 个 labeled samples
• 每个 sample 都会比较 classifier-vs-fallback 的 winner
• 还会显式报告 missing tags 和 extra tags
• 不只看 exact match，也会看 average match score

1. assist mode 现在有 confidence gate

• 公开配置里加入了 minimum_confidence = 0.6
• 低 confidence 的 classifier 输出仍然可以在 calibration / shadow mode 里看到
• 但不会再直接混进 assist-mode enrichment

1. benchmark 和 signal foundation 继续保持稳定

• memory_expected_top1_accuracy = 1.0
• file_scan_expected_top1_accuracy = 0.636
• signal lifecycle 仍然是 claim -> extend -> ack / expire / reclaim

当前 calibration snapshot：

• classifier_exact_match_rate = 0.9
• fallback_exact_match_rate = 0.0
• classifier_better_count = 9
• fallback_better_count = 1
• classifier_filtered_low_confidence_count = 1
• 78 passed

这版真正想做的是：

• 让 assist rollout 受 reviewed evidence 约束
• 不让低 confidence 的模型输出悄悄改写 durable memory

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

Release:
https://github.com/zzhang82/Agent-Memory-Bridge/releases/tag/v0.6.3