Commit Graph
7 Commits
Author SHA1 Message Date
小唯 cb450e9c9d feat(memoryweave): GA-P0-3 拆分 is_deleted 删除原因 → delete_reason 枚举(禁部署)
- 新增 delete_reason 固定枚举列(merge_dup/contradiction/manual/retention/unknown)
- LanceDB(Rust arrow_schema+MemoryRecord+read/insert) 与 sqlite 双侧 additive 同步
- 所有软删写入点统一归一化写 delete_reason:
  SoftDelete(manual/merge_dup/contradiction/retention) · SoftDeleteBatch(unknown)
  · PurgeInvalid(retention) · Rust soft_delete_batch(unknown)/purge_invalid(retention)
- 新增 lancedb_scan_deleted IPC + Go ScanDeleted:存量已删行回填工具枚举用
- 回填工具 go/cmd/backfill_deletereason(dry-run 默认)+ 单测 7 例(每 reason≥1 写入用例)
- 直方图(副盘真跑):N=11240 全 unknown(tombstones 无 provenance,按 AC 判不出写 unknown)
- 禁部署/禁真删/禁改已删记录其它字段:live 库零改动
2026-10-12 18:52:09 +08:00
小唯 28f6951061 feat(memoryweave): GA-P0-2 拆分 recall_count → recall_seen_count / recall_verified_count(禁部署)
背景:recall_count 被当成 V 值喂给衰减/遗忘逻辑,但每次召回命中都 ++,
造成「被召回 500+ 次所以永不遗忘」的虚高保命。实测 11576 条 recall_count>0、
2732 条 >=10,全是「出现在结果里」而非「被真正用上」。t_268b0353 注释早已
点明「被召回 ≠ 被验证」。

改法(AC1-6,禁部署):
- LanceDB(arrow_schema) + sqlite(create/alter) 两侧同步新增
  recall_seen_count / recall_verified_count(additive,不删旧列 recall_count)
- 拆分两个写入点:召回命中(缓冲 Record→UpdateRecallBatch / 兜底 incrementRecallCount)
  → 写 recall_seen_count;正反馈(IncrementUseful)→ 写 recall_verified_count + useful_count
- recall_count 冻结不再增长,保留只读(deprecated),衰减/遗忘只认 recall_verified_count
- 排序去偏(lancedb_ipc Search importance)改用 recall_seen_count(观测/选择信号);
  技能候选筛选(query_memories / GetSkillCandidates)改用 recall_seen_count
- 单测 TestRecallSeenVerifiedIsolation / TestRecallCountFrozen:一次召回 seen++
  verified 不增;一次正反馈 verified++;recall_count 冻结
- 实网取证:副本 LanceDB(禁部署)上跑真实 recall+反馈,
  BEFORE(0/0/0) → AFTER_RECALL(recall_count=0 seen=1 verified=0) →
  AFTER_VERIFY(0/1/1);schema 演进在副本上补列 4 个成功

go build/vet/test + rust build 全绿;E2E 蒸馏触发测试系既有 flake(stash 复现),
本改动无关。禁部署遵守:live /var/lib/memoryweave 未动,仅在副本上验过。
2026-10-12 16:54:32 +08:00
小唯 9b7b969cd3 feat(memoryweave): GA-P0-1 拆分 not_useful_count → human_not_useful_count / dedup_downweight_count
背景:not_useful_count 被两种语义混写——用户人工点「没用」(feedback.go
MarkNotUseful) 与 dedup/合并流程自动降权(core.go Dedup demote) 都写同一字段。
实测 35010 行里 509 条非零但 human useful 仅 21 ⇒ 字段被混写。P2 自适应阈值
要靠它当标签,统计即垃圾。

改法(AC1-5,禁部署):
- LanceDB(arrow_schema) + sqlite(create/alter) 两侧同步新增
  human_not_useful_count / dedup_downweight_count(additive,不删旧列)
- 两个写入点各写各的:feedback.go→IncrementHumanNotUseful;
  core.go dedup→IncrementDedupDownweight
- not_useful_count 保留 deprecated,新代码不再读(读侧仅 Scan/迁移保留)
- 存量迁移工具 cmd/split_notuseful(--dry-run 默认,apply 需终审);
  实测存量无可按来源判定行(M=0)→ 全部归 unknown(K=109 存活行),apply=no-op
- 单测 TestSplitNotUsefulFieldIsolation:两写入点写不同字段,相对增量断言

go build/vet/test + rust build 全绿;rust open() schema 演进在副本上
实测补列 human_not_useful_count/dedup_downweight_count 成功。
2026-10-12 14:17:31 +08:00
小唯 9bfe4c217f fix(p1-1): forget path - counting test + purge-invalid endpoint (t_37ac8a45) 2026-10-08 15:44:04 +08:00
小唯 b5f59703ab fix(storage): check error returns from MemLanceClient in SQLiteMemClient
SQLiteMemClient.IncrementUseful/IncrementNotUseful were discarding the
error return of MemLanceClient.IncrementUseful/IncrementNotUseful.
Since the latter now return real errors (6e67f3e), dropping them
re-introduces the silent-failure class of bug we just fixed across
the Rust path. Also propagate persistMemory errors.
2026-09-20 11:39:30 +08:00
小唯 6e67f3e33f fix(storage): fix all no-op/fake implementations in LanceDB interfaces
- IncrementUseful/IncrementNotUseful now return error and persist via
  rc.Update() (Rust sidecar); all 4 implementations updated
- SoftDeleteBatch now delegates to SoftDelete() for real LanceDB write
- UpdateMemoryContent now persists content + version + version history
- Backup changed from return nil to explicit error (Rust sidecar unsupported)
- IncrementUseful/IncrementNotUseful signatures now include error return;
  all 3 call sites (feedback.go, core.go) check the error
- Test mocks (mockLance, fakeLDB) updated to match new interface

Fixes the 'silent no-op' class of bugs discovered during dedup repair.
2026-09-20 09:53:00 +08:00
xiaowei 2a350cf781 fix(distill): max_tokens 800→1200 — Agnes推理token预算不足
Agnes 2.0-flash是推理模型,推理token占大量预算。
800仍不够→JSON截断(parse error: unexpected end of JSON input)。
1200确保推理+正文都够用。
2026-08-20 01:13:06 +08:00