feat: improve memory data quality

[lishuceo]

Summary

• 提取质量优化: 在 extractor prompt 中增加反例规则（拒绝 PR 状态、修复方案、部署信息等临时事务），收紧类型定义（decision 仅用于架构决策），增加代码层面校验（最短长度、事务性模式过滤、无 TTL 的 state 降级为 fact）
• 置信度自然升级: evidence_count >= 3 时自动将 L0 提升为 L1（confidence 上限从 0.7 → 0.9），无需 UI 交互
• 存量数据治理: 一次性迁移脚本归一化碎片化的 workspace_dir（5 种本地路径 → 1 个规范标识），批量失效约 80 条事务性记忆
• 超替阈值调优: 相似度阈值从 0.85 降到 0.78，提升语义相近记忆的合并率

Test plan

• 3 个测试文件新增 12 个测试用例，覆盖所有新逻辑
• 全量测试 1123/1124 通过（1 个预期的 quality.test.ts 向量搜索质量测试失败，与本次改动无关）

🤖 Generated with Claude Code

[claude]

Claude finished @lishuceo's task in 1m 26s —— View job

---

PR Review: Memory Data Quality Improvements

• Check for previous review comments
• Read changed files and understand context
• Review code for bugs, security, and architecture issues
• Post review summary

---

⚠️ Issues Found

Severity	File	Issue	Confidence
🟡 Warning	src/memory/store.ts:173	Stale JSDoc — says "similarity > 0.85 / distance < 0.15" but threshold was changed to 0.78/0.22	90

Overall Assessment

Solid PR. The changes are well-structured and follow existing patterns:

• Extraction quality: The expanded prompt with anti-examples and the code-level guards (TRANSIENT_PATTERNS, MIN_CONTENT_LENGTH, state→fact demotion) form a good defense-in-depth — the LLM prompt steers extraction, and validateMemories() catches what leaks through.
• Auto-promotion: The L0→L1 promotion logic in updateEvidence() is clean — reads after the synchronous db.updateEvidence() call, so no race conditions with better-sqlite3's synchronous API.
• Data migrations: The PRAGMA user_version guard is the right approach. Migration is wrapped in a transaction with idempotent checks. Test coverage for the "should not re-run migrations on subsequent opens" case is a nice touch.
• Similarity threshold: Lowering from 0.85→0.78 is a reasonable tuning — just update the JSDoc to match (see inline comment).
• Tests: 12 new test cases cover all new logic paths well. Test content was updated to be ≥15 chars to match the new MIN_CONTENT_LENGTH, which is good attention to detail.

[claude]

🟡 Stale JSDoc (confidence: 90): The JSDoc on line 173 still says "similarity > 0.85 (distance < 0.15)" but this line changed the threshold to 0.78/0.22. Please update the doc to match.

Fix this →

[claude]

⚠️ Still exists. The inline comment on line 164 was updated correctly (cosine distance < 0.22 ≈ similarity > 0.78), but the JSDoc on line 147 still reads similarity > 0.85 (distance < 0.15 for cosine). Please update the JSDoc to match.

[claude]

⚠️ Still unfixed after 3 reviews. Line 147 JSDoc says similarity > 0.85 (distance < 0.15 for cosine) but the threshold on line 164-165 is 0.78/0.22.

[claude]

Claude finished @lishuceo's task in 2m 3s —— View job

---

PR Review: Memory Data Quality Improvements (Re-review)

• Check previous review comments and resolve fixed issues
• Read changed files and review new commits
• Post review summary

---

⚠️ Issues Found

Severity	File	Issue	Confidence
🟡 Warning	src/memory/store.ts:147	Stale JSDoc (still unfixed): JSDoc says "similarity > 0.85 (distance < 0.15 for cosine)" but the threshold was changed to 0.78/0.22 on line 164-165	90

1 previous issue persists — replied on the existing thread.

Overall Assessment

The two new commits (e6fe310 simplify + cd501b8 test updates) are clean. The supersedes/supersede_reason column removal and confidence-from-scoring removal are well-executed:

• Column removal is safe — existing DBs keep the unused columns silently; fresh DBs never create them. SELECT * in getMemory() returns whatever columns exist, and rowToMemory() no longer maps them, so no breakage either way.
• _store: unknown parameter in createMemorySearchMcpServer maintains API compatibility with the caller in executor.ts:422 while correctly signaling the param is unused.
• Score simplification (search.ts:142) — removing * memory.confidence from finalScore is consistent with the broader theme of dropping confidence as a ranking signal.
• Migration guard — PRAGMA user_version is the right approach. Existing DBs get user_version=0 by default, so the v1 migration runs exactly once on upgrade. The transaction wrapping is correct.
• Tests updated — all test files properly adapted to the simplified API (no supersedes, no supersedeReason, no getSupersedChain).

Only action item: update the JSDoc on store.ts:147 from 0.85/0.15 to 0.78/0.22.

[claude]

Claude finished @lishuceo's task in 1m 39s —— View job

---

PR Review: Memory Data Quality Improvements (3rd review)

• Check previous review comments and resolve fixed issues
• Review current diff for new issues
• Post review summary

---

⚠️ Issues Found

Severity	File	Issue	Confidence
🟡 Warning	src/memory/store.ts:147	Stale JSDoc (3rd review, still unfixed): JSDoc says similarity > 0.85 (distance < 0.15 for cosine) but threshold is 0.78/0.22 on lines 164-165	90

0 previous issues fixed — the stale JSDoc from review #1 persists. Replied on the existing thread.

Overall Assessment

The PR is solid. This review covers 5 commits across 29 files (including the skill-creator/nano-pdf additions from the base branch merge).

Memory changes (core of this PR):

• Extraction quality guards are well-layered: prompt anti-examples + TRANSIENT_PATTERNS regex + MIN_CONTENT_LENGTH + state→fact demotion. Defense-in-depth is the right approach.
• Supersede chain removal is clean — supersedes/supersede_reason columns dropped from MemoryRow, Memory, MemoryCreateInput, and all downstream code. Existing DBs with these columns are unaffected (SQLite ignores extra columns on SELECT).
• Data migrations via PRAGMA user_version are correctly guarded and wrapped in a transaction. The workspace normalization and transient memory invalidation are idempotent.
• Similarity threshold (0.85→0.78) and score simplification (removing * memory.confidence) are consistent with the broader theme.

Document support (new in this sync):

• DocumentAttachment type, downloadMessageFile(), and the file parsing in parseMessage() are correctly wired through the entire pipeline: parseMessage → taskQueue.enqueue → processQueue → executeClaudeTask/executeDirectTask → ClaudeExecutor. All call sites match function signatures.
• 30MB file size limit and PDF-only restriction are sensible guards.
• buildMultimodalPrompt correctly handles the new documents parameter with proper content block ordering (documents → images → text).

Only action item: Update the JSDoc on store.ts:147 from 0.85/0.15 to 0.78/0.22.