v5.1.0 — Production Hardening + CI Enforcement

Fixed (end-to-end runtime)

• save_message 端到端契约: 修复 conversations 行从未创建导致 search_messages INNER JOIN 返回 0 行——命名空间隔离从此从未生效
• seq 字段自动分配: 客户端不传 seq 时自动分配，避免 NOT NULL 约束失败
• 6 个端到端集成测试: test_mcp_e2e.py 真实 spawn MCP server 子进程验证 round-trip
• DeprecationWarning 消除: datetime.utcnow() → datetime.now(timezone.utc)，全部 0 警告
• semantic_memory.py 导入路径: from prompts → from src.prompts
• database.py __del__: 自动关闭 sqlite 连接，消除 ResourceWarning
• TypeScript TS2742 修复: definePluginEntry() 默认导出加显式类型注解

Changed (docs)

• README.md: 移除未核实的 arXiv ID，MCP 工具数 22 → 17
• mcp_server/README.md: 整段重写，列出 v5.0 实际 17 个工具
• 6 个 docs 加 v2-era 弃用 banner

Removed (dead code)

• scripts/ (20 个文件, ~200 KB): v2.x 时代脚本，功能已迁入 src/pipeline/
• systemd/ (6 个文件): ExecStart 指向不存在的脚本

Changed (lint + CI)

• Black 格式化: 78 个 Python 文件统一格式
• autoflake: 移除 87 个未使用 import，51 个空 f-string
• .flake8 配置: 忽略 E203/E402/E501/F841（Black 冲突 + 结构性）
• continue-on-error 全部移除: CI 失败现在真正阻断 PR
• npm install 修复: 移除 | head -100 管道截断

Dependencies

• openclaw 2026.5.27 → 2026.6.6
• vitest 4.1.7 → 4.1.8
• @vitest/coverage-v8 4.1.7 → 4.1.8
• @types/node 25.9.1 → 25.9.3

Stats

• Tests: 634 passed, 0 failed, 4 skipped
• Coverage: 77% (gate: 60%)
• Security alerts: 0 open
• Lint: Black ✅ isort ✅ flake8 ✅ tsc ✅

Full Changelog: v5.0.3...v5.1.0

🔧 Engineering Hardening

Comprehensive code quality and security improvements across the entire codebase.

P0 — Must-fix

P1 — Code Quality

Stats

• Net code reduction: ~400 lines
• Tests: 481 passed, 0 failed
• Security: 0 vulnerabilities (npm audit)

Upgrade

git pull origin master
pip install -e .   # Now works with pyproject.toml

Full Changelog: v5.0.2...v5.0.3

Fixed (OpenClaw v2026.4.2 Compatibility)

• 🐛 definePluginEntry(): 迁移到 openclaw/plugin-sdk/plugin-entry 标准导出模式
• 🐛 kind: "context-engine": 在 lobsterEngine.info 中添加类型声明
• 🐛 debugLog 保护: debugLog() 仅在 LOBSTER_DEBUG=1 时写入日志
• 🐛 console.log 删除: register() 末尾的 bare console.log 已移除

Fixed (Config Consistency)

• 🐛 maxContextTokens 统一: schema 默认值和 runtime fallback 均为 40000
• 🐛 过时配置删除: mcp_server/lobster.config.json 已删除

Fixed (Security)

• 🔒 硬编码 API Key → 环境变量
• 🔒 /tmp 路径泄漏修复

v5.0.1 → v5.0.2 — Full Compatibility Patch

This release consolidates all OpenClaw v2026.4.2 compatibility fixes.

v5.0.2 (commit c262c44) — Documentation + P2 Cleanup

P1 - OpenClaw v2026.4.2:

• definePluginEntry() — migrated from plain object export
• kind: "context-engine" — added to lobsterEngine.info
• Hook names verified: before_agent_start, agent_end confirmed in v2026.4.2

P2 - Debug/Logging:

• debugLog() guarded behind LOBSTER_DEBUG=1 env var
• Bare console.log removed

P2 - Config:

• maxContextTokens unified to 40000 everywhere
• mcp_server/lobster.config.json (v1.2.8) deleted

P2 - Security:

• Leaked ZHIPU API key → ZHIPU_API_KEY env var
• Hardcoded /tmp paths → relative project-root paths

Docs:

• CHANGELOG.md: v5.0.2 + v5.0.1 + v5.0.0 entries added
• README_EN.md: v5.0.0 MemOS 4-Phase (full refresh)
• README.md: confirmed v5.0.0 ✅

v5.0.1 (commit 731bce5) — P0 Compatibility Fixes

peerDependencies: >=2026.3.0 → >=2026.4.2
SIGINT/SIGTERM handlers: Python subprocess lifecycle cleanup

P0 Bug Fixes:

• C-HLR+ formula: math.exp(-t/h) → 0.5 ** (t/h)
• Double scoring removed from retention
• Pass 4 lossless: user/assistant skip dedup

Tests: 423 passed

Full Commit History v5.0.x

```
c09f459 docs: update CHANGELOG.md and README_EN.md for v5.0.2
c262c44 fix(v5.0.2): OpenClaw v2026.4.2 compatibility + P2 cleanup
731bce5 fix(v5.0.1): P0 compatibility fixes for OpenClaw v2026.4.2
fb5f378 feat(v5.0): MemOS 4-Phase Optimization
```

v5.0.0 - MemOS 4-Phase Optimization

What's New

Phase 1: Core Intelligence

• Vector Embedder - OpenAI Compatible API + numpy offline fallback, BLOB storage in SQLite
• Hybrid Retriever - FTS5 + Vector → RRF(k=60) → MMR(λ=0.7) → 14d Decay → Filter(≥0.45)
• LLM Fallback Chain - 4-level chain: skill → summary → native → mock
• Dual Decay Parameters - 12h compression / 14d retrieval (independent)

Phase 2: Skill Evolution

• Task Detector - 2h timeout + LLM topic judgment (SAME bias)
• Skill Evolver - Rule filter → LLM eval → SKILL.md → quality score → save
• MCP lobster_skill - get/install/list actions

Phase 3: Multi-Agent

• Owner Field - migrate_v50, owner on messages/summaries/notes
• Public Memory - lobster_memory_write_public
• Skill Sharing - lobster_skill_search (self/public/mix), lobster_skill_publish/unpublish

Phase 4: Engineering

• Viewer Web UI - 127.0.0.1, SHA-256 auth, HttpOnly cookie
• Async Queue - Background worker for embed/task_detect/skill_eval
• OpenClaw Migration - 🦐 prefix, SHA-256 dedup, checkpoint resume

Bug Fixes (P0)

• C-HLR+ formula: math.exp → 0.5 (correct forgetting curve)
• Pass 4 lossless: user/assistant messages skip dedup
• Missing _get_llm() method
• Path traversal vulnerability fix
• Dual scoring removal (access_count bonus)

Test Results

• 436 tests pass (+297 new)
• 22 MCP tools (+7 new)

Breaking Changes

• None - fully backward compatible

🎉 LobsterPress v4.0.0「深海」版本

🚀 核心目标

超越 lossless-claw，实现学术级认知记忆系统

✨ 五大模块重构

🔧 模块一：CMV 三遍无损压缩

• Pass 1: 剥离 base64/长 JSON 冗余
• Pass 2: 去重工具结果
• Pass 3: 折叠系统样板代码
• 无损原则：user/assistant 消息永不修改

🧬 模块二：C-HLR+ 自适应遗忘曲线

• 复杂度驱动半衰期：h = base_h × (1 + α × complexity) × spaced_bonus
• 对数稳定性增长：避免无限膨胀
• 指数底数优化：从 e 改为 2
• Ref: arXiv:2004.11327

⚡ 模块三：Focus 主动压缩触发

• 定时触发：每 12 轮主动压缩
• 紧急触发：上下文使用率 > 85%
• 被动触发：原有阈值逻辑
• Ref: arXiv:2502.15957

🔄 模块四：R³Mem 可逆三层压缩

• Layer 1: Document-Level（返回子摘要）
• Layer 2: Paragraph-Level（返回原始消息）
• Layer 3: Entity-Level（按实体过滤）
• 实体追踪：人名、文件名、概念自动提取
• Ref: arXiv:2502.15957

🛠️ 模块五：WMR 框架重构

• Write 层：lobster_compact + lobster_correct
• Manage 层：lobster_sweep + lobster_assemble + lobster_prune
• Read 层：lobster_grep + lobster_describe + lobster_expand + lobster_status

📚 学术参考

• arXiv:2004.11327 — Adaptive Forgetting Curves for Spaced Repetition
• arXiv:2502.15957 — Focus: Context-Aware Prompt Compression for LLMs
• arXiv:2502.15957 — R³Mem: Bridging Memory Retention and Retrieval
• CMV: Context Maintenance and Retrieval

📦 安装

npm install @sonicbotman/lobster-press@4.0.0

🔗 链接

• GitHub
• npm
• 文档
• 更新日志

Published with ❤️ by 小云 (Xiao Yun)

🧠 v3.0.0 - 认知记忆系统完整版

✨ 新增功能

Feature 3: 语义记忆层 (Semantic Memory Layer)

• 独立于 DAG 的稳定知识库
• LLM 自动提取决策、偏好、约束、事实
• 上下文注入 notes（<500 tokens）
• 去重机制

Feature 4: 矛盾检测与记忆重巩固

• NLI 模型检测（可选）+ 规则降级检测
• 记忆重巩固机制
• 保留完整溯源链

📊 测试结果

4/4 测试通过

🔄 向后兼容

• v2.6.0 schema 完全兼容
• llm_client 参数可选

认知科学驱动: 基于 EM-LLM、HiMem 等研究论文设计

PR: #99
基于: v2.6.0

🎉 Major Release: Lossless Memory System

This release introduces a complete lossless memory architecture with DAG-based hierarchical compression.

✨ New Features

Phase 1: Lossless Storage (100%)

• SQLite + FTS5 full-text search
• Permanent message storage (never lose data)
• DAG structure for traceability

Phase 2: DAG Compression (100%)

• Leaf compression (messages → leaf summaries)
• Condensed compression (summaries → condensed summaries)
• 67.8% compression ratio (416 → 134 tokens)
• Fresh tail protection (last 32 messages)

Phase 3: Agent Tools (100%)

• lobster_grep - Search messages and summaries
• lobster_describe - Inspect summary structures
• lobster_expand - Expand summaries to original messages

Phase 4: Incremental Compression (100%)

• Auto-trigger compression (75% context threshold)
• Real-time monitoring
• Multi-conversation support

📊 Performance

• Compression ratio: 67.8%
• Memory efficiency: 70.8% reduction
• Compression speed: ~2s for 30 messages

🙏 Acknowledgments

Inspired by lossless-claw (Martian Engineering) LCM plugin architecture.

🎉 首次发布

功能特性

• 🦞 智能上下文压缩引擎 v5 - 多维度消息评估，三级压缩策略
• 🧠 消息重要性评估引擎 - 自动识别决策、错误、配置等关键信息
• 🎓 自适应学习系统 - 学习用户偏好，动态调整参数
• 🔮 预测性压缩引擎 - 预测 Token 需求，提前执行压缩
• 💰 成本优化器 - 分析 API 成本，推荐最经济的策略

文档

• 📖 完整的 README（背景、理念、通用性、价值、部署方案）
• 🏗️ 架构设计文档
• 🔌 API 文档
• 🎨 自定义指南
• ❓ FAQ

示例

• 基础用法示例
• 高级配置示例
• 集成示例（Python, Node.js, Bash）

统计

• 核心脚本: 6 个（1289 行代码）
• 文档文件: 4 个（1298 行）
• 示例文件: 3 个
• Systemd 服务: 6 个
• 总大小: 500K

核心价值

• 💰 节省 30-50% Token 成本
• ⚡ 自动化管理，零干预
• 🧠 保留重要信息
• 🎯 个性化体验
• 🔒 本地运行，数据安全

兼容性

• ✅ OpenAI ChatGPT
• ✅ Claude
• ✅ GLM 系列
• ✅ Qwen 系列
• ✅ 任何基于对话的 AI 系统

许可证

MIT License - 自由使用、修改、分发

🦞 让 AI 记忆永不溢出！

Made with ❤️ by the LobsterPress Team