v1.1.0-beta.9

Highlights

• Hybrid retrieval improvements and rerank provider expansion:
  • Add TEI rerank provider support + config schema + docs (PR #207)
  • Add DashScope (Alibaba Cloud) rerank provider (PR #199)
• Temporal facts / supersede semantics for mutable facts + lifecycle tests (PR #183)
• Prompt-injection hardening: sanitize reflection slices/lines before any injection (PR #190)
• Retrieval correctness: propagate excludeInactive across all retrieval paths
• Runtime deps: declare apache-arrow as a direct dependency (PR #194)
• CI/workflow hardening for forks: skip Claude review + PR auto-assign on fork PRs (PR #210)

Install

• npm (beta tag): npm i memory-lancedb-pro@beta

Notes

This is a prerelease build published from the latest master state after 1.1.0-beta.8.

v1.1.0-beta.8

Highlights

• Hybrid retrieval fusion weighting: vectorWeight / bm25Weight now actually affects fusion scoring.
• Docs updates around fusion weighting and scoring defaults.

Merged PRs

• #176 fix(retriever): make vectorWeight/bm25Weight actually affect fusion scoring (v2)
  #176
• #175 revert previous fusion weighting attempt (#168)
  #175
• #168 fusion weighting design + initial implementation (was reverted)
  #168

Notes

• This is a prerelease (beta) for the 1.1.0 line.

v1.1.0-beta.7

Highlights

• Contextual SupportInfo improvements + regression coverage (Support slices, merge/contradict/contextualize decision flows).
• Retrieval correctness fixes: cosine-based vector search scoring + related hybrid/FTS diagnostics.
• Smart metadata v2 parsing/normalization (incl. Chinese alias mapping) with E2E tests.
• Smart extractor noise filtering stability: NoisePrototypeBank degeneracy guard to prevent false-positive noise detection.

Merged PRs

• #164 fix: NoisePrototypeBank degeneracy guard — fix flaky smart-extractor-branches test (by @AliceLJY)
  #164
• #171 Add support info regression coverage (by @chenjiyong)
  #171
• #161 feat: 检索正确性修复 + 情境化 SupportInfo — 适配 master 架构 (by @398618101)
  #161
• #159 fix(recall): remove score/source suffixes from visible recall text (by @Disaster-Terminator)
  #159

Notes

• This is a prerelease (beta) for the 1.1.0 line.

v1.1.0-beta.6

Highlights

• Derived-Focus V2: upgraded historical <derived-focus> ranking (conservative strict/soft normalization, non-linear group scoring, diversity-aware selection).
• Shortlist up to 36 candidates before final injection capped at 13 (no hard score > 0.3 threshold gate).
• Reflection-Recall selection now reuses the shared normalization/aggregation/selection helpers.
• Recall grouping is partitioned by kind + strictKey, keeping invariant/derived rows separate even when normalized text matches.

Notes

• This is a beta release; latest remains unchanged.

Changes

• feat(reflection): upgrade derived-focus ranking and recall selection (#114)

v1.1.0-beta.5 (dev)

Dev channel (smart-memory branch)

This build is published under the npm dist-tag dev.

What’s inside

• Smart-memory feature set from branch dev/smart-memory-v1.1.0 (merged PR #104).
• Includes extended functional/e2e + smart-memory lifecycle regression coverage.

Install

npm i memory-lancedb-pro@dev

Notes

• This dev build may diverge from the mainline beta track.

v1.1.0-beta.3 — Self-Improvement & Memory Reflection (Beta)

🧪 v1.1.0-beta.3 — Self-Improvement & Memory Reflection

This is a beta release. Install with npm install memory-lancedb-pro@beta. The stable latest remains at v1.0.32.

✨ New Features

Self-Improvement Governance (PR #43)

• self_improvement_log tool for structured learning/error/feature-request entries
• self_improvement_review and self_improvement_extract_skill governance tools
• Automatic .learnings/ file bootstrapping on agent startup and session reset

memoryReflection Session Strategy (PR #43)

• New sessionStrategy: "memoryReflection" option for plugin-driven reflection on /new and /reset
• Structured reflection output with invariants, derived focus, and error loop detection
• Optional LanceDB persistence with similarity dedup (> 0.97 threshold)
• Dedicated reflection-agent support via memoryReflection.agentId
• Reflection retry with automatic fallback chain

Markdown Mirror (PR #43)

• Dual-write memory entries to readable Markdown files for audit/debug
• Configurable via mdMirror.enabled and mdMirror.dir

🔧 Fixes

• Embedding chunking config: embedding.chunking toggle now properly honored (PR #87)
• Embedding error diagnostics: Actionable error messages for auth/network/rate-limit failures (PR #67)
• Session strategy compatibility: Legacy sessionMemory.enabled safely maps to new strategy system

📖 Docs

• Added Long Context Chunking (Section 11) and Embedding Error Diagnostics (Section 12) to README
• Clarified recall parameter mapping and practical tuning baselines (PR #77)
• Full EN/CN documentation parity

PRs included

• #43 (feat: memory-reflection + self-improvement)
• #67 (fix: embedding provider failure hints)
• #77 (docs: recall param mapping)
• #87 (fix: honor embedding chunking config)

Install

npm install memory-lancedb-pro@beta

v1.0.32

Highlights

• Strip OpenClaw Conversation info / Sender metadata noise before auto-capture matching.
• Strip the same metadata noise before adaptive retrieval query normalization.
• Parse autoRecallMinRepeated from plugin config so repeated-memory suppression works when configured.

Why this matters

OpenClaw v3.2+ can prepend untrusted metadata blocks to messages. This release prevents those blocks from inflating capture input, skewing adaptive-retrieval heuristics, and causing noisy false-positive recalls.

Fixed

• Auto-capture now evaluates meaningful user content instead of metadata wrappers.
• Adaptive retrieval now removes metadata blocks consistently before length/trigger logic runs.
• autoRecallMinRepeated is now read from plugin config during parsing.

Breaking changes

• None. Backward compatible.

References

• PR #50: #50
• Merge commit: 6ffb5bc

Verification

• npm install
• npm test
• OpenClaw CLI smoke suite passed locally before release.

Thanks @kctony for the metadata-noise fix.

v1.0.31

Fixed

• memory-pro import now preserves provided IDs and is idempotent: if an imported entry has an id that already exists in the target scope/db, it will be skipped instead of creating duplicates.

Notes

• Import without an id keeps the previous best-effort similarity dedupe behavior.

v1.0.30

Highlights

• Make the dedup pre-check fail-open so it never blocks legitimate memory writes.
  • Auto-capture: if vectorSearch() fails, we log a warning and continue storing.
  • memory_store tool: if vectorSearch() fails, we warn and continue storing.

PRs

• #45 fix: fail-open dedup pre-check to avoid blocking memory writes (@marcustseng-agent)

v1.0.29

Highlights

• Improve normalizeQuery() noise stripping: removes OpenClaw v3.2 injected metadata headers (e.g. Conversation info (untrusted metadata): / Sender (untrusted metadata):) and trims after the JSON block.
• Keeps existing stripping for [cron:...] wrapper prefixes and timestamp prefixes.

PRs

• #47 fix: enhance normalizeQuery to strip OpenClaw v3.2 metadata and timestamps (@kctony)