Closed
Conversation
…-rerank fix: add TEI rerank provider support
…view-on-fork-prs ci: skip Claude Code Review on fork PRs
…ortexReach#221) fix(cli): flush pending writes after import by calling store.close()
…close() (CortexReach#221)" (CortexReach#232) revert: undo CortexReach#221 - store.close() based on incorrect root cause
feat: skip USER.md-exclusive facts in plugin memory
…nostics Improve LLM diagnostics and make timeouts configurable
When a large CJK text (14KB+ Chinese .md file) is processed by auto-recall, embedSingle() enters an infinite recursion loop because: 1. smartChunk() treats token limits as character limits, but CJK characters use 2-3x more tokens than ASCII characters 2. Chunks of 5740 chars (70% of 8192 token limit) still exceed the model's token context for CJK text 3. smartChunk() returns 1 chunk identical to input → embedSingle() recurses with the same text → infinite loop This produced ~50,000 embedding errors in 12 minutes, blocking the entire Node.js event loop and making all agents unresponsive. Fixes: - Add recursion depth limit (max 3) to embedSingle() with forced truncation as fallback - Detect single-chunk output (same size as input) and truncate instead of recursing - Add CJK-aware chunk sizing in smartChunk() (divide char limit by 2.5 when CJK ratio > 30%) - Truncate auto-recall query to 1000 chars before embedding - Add 10s global timeout on embedPassage()/embedQuery() Closes CortexReach#214 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…e (PR CortexReach#215 follow-up) This commit addresses the two blocking issues raised in PR CortexReach#215: 1. Timeout now uses AbortController for TRUE request cancellation - Timer is properly cleaned up in .finally() - AbortSignal is passed through to embedWithRetry 2. Recursion now guarantees monotonic convergence - Introduced STRICT_REDUCTION_FACTOR = 0.5 - Each recursion level must reduce input by 50% - Works regardless of model context size Modified by AI assistant (not human code) based on PR CortexReach#215. Thanks to the original author and maintainers. Co-authored-by: Hi-Jiajun <Hi-Jiajun@users.noreply.github.com>
… section - Highlight "AI Memory Assistant" as the core value proposition with before/after demo - Add Ecosystem section featuring setup script (CortexReach/toolbox) and Claude Code/OpenClaw skill - Move comparison table into collapsible details - Fix RRF → Hybrid Fusion naming to match actual implementation - Add reflection to storage categories (aligns with store.ts/tools.ts) - Clarify 6-category semantic labels vs storage categories in schema docs - Add openclaw plugins install as primary manual install path - Fix minor English expressions and Chinese translation polish - Reviewed by CC (Claude Code) + Codex across 3 rounds Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…eedback (CortexReach#216) Voyage AI (and Voyage-style proxies) expose an OpenAI-compatible embeddings endpoint but reject standard OpenAI request fields and use different field names for task hints and output dimensions. Without handling these differences, all embedding calls to Voyage endpoints failed with 400 Bad Request: Unknown request body key: encoding_format, and task/dimension configuration was silently dropped. Changes: - Introduce EmbeddingCapabilities with taskField/taskValueMap/ dimensionsField replacing boolean flags; buildPayload is now fully data-driven with no provider conditionals - voyage-compatible profile: maps task values (retrieval.query→query, retrieval.passage→document) and sends output_dimension instead of dimensions; suppresses encoding_format - Remove unused _profile field from Embedder (capabilities already cached via _capabilities) - Add debug warnings in Embedder constructor when normalized or taskQuery/taskPassage are configured but the provider profile does not support them, so misconfiguration surfaces immediately - Fix Jina-specific auth hint to fire for jina-* models behind a proxy by checking profile instead of provider label; remove dead code - Add voyage-3/3-lite/code-3/3-large to EMBEDDING_DIMENSIONS - Add Voyage row to README embedding providers table - Add tests: Voyage input_type translation, output_dimension field, constructor debug warnings, Jina proxy auth hint; extract captureDebug helper to reduce warning test boilerplate Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…on (CortexReach#226) 核心变更： - ACL 键规范化：trim agentAccess 键，避免空白填充导致 ACL 失效 - 明确 deny-all 语义：scopeFilter=[] 拒绝所有读写，与 undefined（绕过）区分 - 统一 hook agentId 解析：before_agent_start 和 before_prompt_build 均使用 resolveHookAgentId - importConfig 原子性：验证失败时警告不泄漏，配置完整回滚 - 保留 bypass ID 限制：system/undefined 仅限内部使用，拒绝配置和 sessionKey 解析 - reflection 加载安全：bypass 调用方使用可选 scopeFilter，避免过滤绕过 新增测试： - test/reflection-bypass-hook.test.mjs - hook bypass 和 sessionKey 解析 - test/scope-access-undefined.test.mjs - bypass ID 拒绝和 scope 访问 - test/smart-extractor-scope-filter.test.mjs - SmartExtractor scopeFilter 语义 - test/store-empty-scope-filter.test.mjs - 空数组 deny-all 语义 影响范围：6 个源文件 + 4 个测试文件，约 970 行净变动
…rtexReach#217) * feat: add provider-aware OAuth login for memory-pro * fix: address oauth review feedback * fix: preserve llm baseURL and correct oauth callback/path handling * fix(cli): restore pre-oauth llm config on logout * fix(cli): tighten oauth path and logout restore --------- Co-authored-by: Heng Xia <pope@Hengs-Mac-mini.local>
- Remove unused SAFE_CHAR_LIMITS, getSafeCharLimit, DEFAULT_SAFE_CHAR_LIMIT - Add comment explaining batch timeout asymmetry (embedBatchQuery/embedBatchPassage not wrapped) - Note: withTimeout already has .finally() cleanup, no change needed
…ortexReach#238) - Test single-chunk detection (force-reduce when chunk >= 90% of original) - Test depth limit termination (depth >= MAX_EMBED_DEPTH throws) - Test CJK-aware chunk sizing (>30% CJK -> smaller chunks) - Test strict reduction factor (50% per recursion level) - Test batch embedding works correctly
- Preserve and surface chunkError instead of hiding behind original error - Remove 1000 char hard floor in smartChunk for small-context models (now 200) - Add regression test for small-context model chunking (all-MiniLM-L6-v2) - Add regression test for chunkError preservation - Wire cjk-recursion-regression.test.mjs into main test suite (CI)
Fix CI version-sync check failure (package.json and openclaw.plugin.json must have matching versions). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Default raised from 3s to 5s for remote embedding APIs behind proxies.
Configurable in openclaw.json:
"plugins": {
"entries": {
"memory-lancedb-pro": {
"config": {
"autoRecallTimeoutMs": 8000
}
}
}
}
Closes CortexReach#314
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…e link Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…fined agentId (CortexReach#288, CortexReach#231) Root cause: isAccessible(scope, undefined) used bypass mode (any valid scope returns true), but getScopeFilter(undefined) returned only explicitly defined scopes (["global"]), excluding implicit agent scopes like "agent:main". This caused memory_update to reject accessible memories while recall/forget worked. Changes: - scopes.ts: getScopeFilter now returns undefined (full bypass) when agentId is missing, matching isAccessible's existing bypass behavior - tools.ts: memory_update uses runtimeContext.scopeManager instead of context.scopeManager, consistent with recall/forget - test: updated scope-access-undefined assertions to match new bypass semantics Note: beta.10's resolveRuntimeAgentId already ensures agentId defaults to "main" in normal operation. This fix is defensive — prevents the same bug from recurring if any future code path passes undefined agentId. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
… columns (CortexReach#326) Previously doInitialize() only logged a warning when detecting missing columns but never actually added them, causing stats/list/search/update/delete to fail with "No field named scope" on databases upgraded from legacy memory-lancedb. Now uses table.schema() to detect missing columns and table.addColumns() to add them with sensible defaults (scope='global', timestamp=0.0, metadata='{}'). Handles concurrent initialization race via "already exists" error detection. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add recallMode config (full/summary/adaptive/off) for granular control over auto-recall injection depth. Adaptive mode uses zero-LLM-cost pattern matching to analyze query intent and route to appropriate memory categories with depth-aware formatting. - New: src/intent-analyzer.ts — analyzeIntent(), applyCategoryBoost() - New: test/intent-analyzer.test.mjs — 23 tests (all passing) - Modified: index.ts — integrate intent analysis into before_prompt_build - Modified: openclaw.plugin.json — add recallMode schema + UI field Rewritten for v1.1.0-beta.10 (replaces PR CortexReach#313 which was based on beta.9). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: yjjheizhu <yjjheizhu@users.noreply.github.com>
…etadata before extraction
feat: add recallMode with adaptive intent routing (beta.10)
Adds a `MemoryCompactor` that periodically consolidates semantically
similar old memories into single refined entries, inspired by the
progressive summarization pattern (MemOS). Over time, related memory
fragments are merged rather than accumulated, reducing retrieval noise
and keeping the LanceDB index lean.
Key additions:
- `src/memory-compactor.ts`: pure, dependency-free compaction module
with cosine-similarity clustering, greedy seed expansion, and
rule-based merge (dedup lines, max importance, plurality category)
- `store.ts`: new `fetchForCompaction()` method that fetches old entries
with vectors (intentionally omitted from `list()` for performance)
- `index.ts`: `memory_compact` management tool (requires
`enableManagementTools: true`) + optional auto-compaction at
`gateway_start` with configurable cooldown
- `openclaw.plugin.json`: `memoryCompaction` config schema + uiHints
- `test/memory-compactor.test.mjs`: 23 tests, 100% pass
Config example:
memoryCompaction:
enabled: true # auto-run at gateway_start
minAgeDays: 7 # only touch memories ≥ 7 days old
similarityThreshold: 0.88
cooldownHours: 24
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
feat: memory compaction — progressive summarization for stored memories
Combines the following rebased changes onto latest cortexreach/master: - feat: add session compression and adaptive extraction throttling - fix: disable extraction throttle and session compression in smart-extractor-branches test - fix: address Codex review for session compression - fix: address PR CortexReach#318 review — align opt-in defaults, CJK-aware scoring, rate limiter docs Session compression scores conversation texts before auto-capture and drops low-signal content; extraction throttling adds a sliding-window rate limiter and a low-value conversation skip gate. Both features default to opt-in (disabled) to avoid surprising existing users. Resolves merge conflicts with memory-compaction (CortexReach#343) and recall-mode-v2 (CortexReach#342) that landed on master since the original branch.
…sion feat: session compression + adaptive extraction throttling
* feat: add observable retrieval traces, stats collector, and batch dedup Observable Retrieval (Feature 3): - Add TraceCollector (src/retrieval-trace.ts) that tracks entry IDs through each pipeline stage, computing drops, score ranges, and timing - Add RetrievalStatsCollector (src/retrieval-stats.ts) for aggregate query metrics: latency percentiles, zero-result rate, top drop stages - Instrument all retrieval stages in MemoryRetriever.retrieve() with optional trace (zero overhead when disabled via optional chaining) - Add retrieveWithTrace() for always-on debug tracing - Add memory_debug tool (requires enableManagementTools) returning full per-stage pipeline trace with drop info - Extend memory_stats tool to include retrieval quality metrics Batch Dedup (Feature 2): - Add batchDedup() (src/batch-dedup.ts) for cosine similarity dedup within extraction batches before expensive LLM dedup calls - Add ExtractionCostStats tracking (batchDeduped, durationMs, llmCalls) Tests: 26 new tests (16 trace + 10 batch-dedup), all passing. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: address Codex review for observable retrieval 1. [High] Fix hybrid trace timing: replace sequential vector_search/bm25_search stages with single parallel_search stage that correctly represents concurrent execution 2. [High] Fix negative drop display: search stages with input=0 now show "found N" instead of "dropped -N" 3. [Medium] Fix rerankUsed overcount: only emit rerank trace stage when rerank is actually enabled (config.rerank !== "none"), not on every query Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: correct trace mode for BM25 tag queries + batchDedup TODO - RetrievalTrace.mode now includes "bm25" variant - retrieve() and retrieveWithTrace() record mode as "bm25" when tagTokens trigger bm25OnlyRetrieval (was incorrectly "vector") - Add TODO for wiring batchDedup() into extraction pipeline Addresses review feedback from rwmjhb on PR CortexReach#319. * fix: address PR CortexReach#319 review feedback (rwmjhb) 1. Wire batchDedup into extraction pipeline (smart-extractor.ts) - Import and call batchDedup() between LLM extraction and per-candidate dedup, embedding candidate abstracts upfront and filtering near-dupes before expensive LLM dedup calls - Graceful fallback: if embedding or dedup fails, all candidates proceed - Batch-deduped candidates counted in stats.skipped 2. Fix memory_debug scope parameter bug (tools.ts) - resolveScopeFilter() only accepts 2 params (scopeManager, agentId), not 3 — was passing `scope` as a third arg which was silently ignored - Match the pattern used by all other tools: resolve default scope filter first, then override with [scope] if scope param is provided and accessible --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: VACInc <3279061+VACInc@users.noreply.github.com>
…L0 (CortexReach#334) * fix: memory_recall with includeFullText returns L2 content instead of L0 When an agent explicitly calls memory_recall with includeFullText: true, it signals active intent to retrieve detailed memory content. Previously this still returned r.entry.text (which stores the L0 abstract), making includeFullText a no-op. Fix: resolve full content via L2 → L1 → L0 fallback chain so agents get the complete narrative when they ask for it. Also expose fullText in sanitizeMemoryForSerialization so the details.memories payload carries L2 content for callers that inspect the structured response. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * test: add regression test for memory_recall includeFullText L2 content Verify that passing includeFullText:true to memory_recall returns L2 content (not the L0 abstract) in both the rendered text output and details.memories[].fullText. Also asserts that details.memories[].text still carries L0 for backwards compatibility. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix: scope fullText to memory_recall handler, add opt-in and legacy tests Move fullText injection out of sanitizeMemoryForSerialization (shared helper used by all tools) and into the memory_recall execute handler, injected only when includeFullText=true. This keeps the opt-in contract clean and avoids L2 content bloating responses from other tools. Add two regression tests: - includeFullText=false: details.memories[].fullText is undefined - legacy memory (no smart metadata): fullText falls back to entry.text Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: linjiaqiang <linjiaqiang@maitrust.com> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…ion cap + ephemeral marking (CortexReach#353) Fixes CortexReach#345 — autoRecall could cause context to balloon and trigger compaction loops due to three compounding issues: 1. Session state leak: recallHistory and turnCounter Maps were never cleaned on session_end, causing unbounded growth across sessions. Added session_end handler to clean up auto-recall Maps. 2. No hard per-turn safety cap: Added maxRecallPerTurn config (default 10) that acts as a ceiling on top of autoRecallMaxItems to prevent misconfiguration from causing context inflation. 3. Injection blocks persisted in transcript: Added ephemeral: true flag to the returned hook result so the host framework's compaction logic can safely discard injected <relevant-memories> blocks instead of persisting them into the session transcript. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ll (CortexReach#354) Fixes CortexReach#350 — autoCapture was writing state: "pending" in metadata but autoRecall governance filter requires state === "confirmed", creating a deadlock where auto-captured memories could never be auto-recalled. Changed all auto-capture write paths (regex fallback in index.ts and all 4 Smart Extraction paths in src/smart-extractor.ts) to write state: "confirmed" directly. This unblocks the core autoCapture + autoRecall workflow without requiring users to enable management tools and manually promote every memory. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
CortexReach#391) * fix: default selfImprovement.enabled to true when config block omitted (CortexReach#375) When the selfImprovement config block is entirely omitted from plugin config, enabled defaults to false, preventing self_improvement_* tools from registering. This is inconsistent with the behavior when the block IS provided but enabled is not set, which correctly defaults to true via the !== false check. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: align selfImprovement sibling flags defaults when config block omitted When the selfImprovement config block is entirely omitted, beforeResetNote, skipSubagentBootstrap, and ensureLearningFiles now default to true — matching the behavior when an empty selfImprovement object is provided. Addresses review feedback on CortexReach#391. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…works (CortexReach#383) * fix: use native fetch for Ollama embedding to ensure AbortController works Root cause: OpenAI SDK HTTP client does not reliably abort Ollama TCP connections when AbortController.abort() fires in Node.js. This causes stalled sockets that hang until the gateway-level 120s timeout. Fix: Add isOllamaProvider() to detect localhost:11434 endpoints, and embedWithNativeFetch() using Node.js 18+ native fetch instead of the OpenAI SDK. Native fetch properly closes TCP connections on abort. Also adds Test 8 to cjk-recursion-regression.test.mjs with assertion updated to match Embedding provider unreachable error (AliceLJY review fix). Fixes CortexReach#361. * fix(embedder): thread external AbortSignal + fix testOllamaAbort regression --------- Co-authored-by: James <james53882@users.noreply.github.com>
…ortexReach#387) * fix: suppress plugin registration logs in CLI mode (CortexReach#380) In CLI mode (OPENCLAW_CLI=1), registration and lifecycle logs are downgraded from info to debug to avoid flooding stderr with ~11 lines of noise before command output. Affected logs: - smart extraction enabled - plugin registered (db, model, smartExtraction) - diagnostic build tag loaded - session-strategy: using none - session-memory: typed before_reset hook registered Gateway runtime behavior is unchanged (logs remain at info level). * fix: suppress self-improvement and memory-reflection registration logs in CLI mode --------- Co-authored-by: 钟远 <15448329+zhong-yuan9527@user.noreply.gitee.com>
Add `memory-pro import-markdown` command to migrate existing Markdown memories (MEMORY.md, memory/YYYY-MM-DD.md) into the plugin LanceDB store for semantic recall. This addresses Issue CortexReach#344 by providing a migration path from the Markdown layer to the plugin memory layer.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
20 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Add memory-pro import-markdown CLI command to migrate existing Markdown memories (MEMORY.md, memory/YYYY-MM-DD.md) into the plugin LanceDB store for semantic recall.
This addresses Issue #344 by providing a migration path from the Markdown layer to the plugin memory layer.
Changes
Usage
\\�ash
npx memory-lancedb-pro memory-pro import-markdown
\\
Related