Releases: Asher-2000/dsh-memory-connect
Releases · Asher-2000/dsh-memory-connect
Release list
v0.2.0 — Context Explosion Prevention
🛡️ Context Explosion Prevention
New Features
- TokenCounter — Token estimation for English, Chinese, and mixed text
- Smart Prioritization — Memory ranking by relevance × recency × frequency
- Budget Management — Configurable
maxContextTokenslimit (default: 4000) - Memory Compression — Auto-truncation when approaching token limits
Configuration
maxContextTokens: 4000 # Max tokens for memory injection
reservedTokens: 8000 # Tokens reserved for other context
smartPrioritization: true # Enable smart prioritization
enableCompression: true # Enable compressionImprovements
- Updated README with context explosion prevention docs
- Added context budget tests
- Improved memory selection algorithm
🛡️ 上下文爆炸防护
新功能
- Token 计数器 — 支持中英文混合 token 估算
- 智能优先级 — 基于相关性 × 时间 × 频率的记忆排序
- 预算管理 — 可配置
maxContextTokens限制(默认 4000) - 记忆压缩 — 接近限制时自动截断
配置
maxContextTokens: 4000 # 记忆注入的最大 token 数
reservedTokens: 8000 # 为其他上下文预留的 token
smartPrioritization: true # 启用智能优先级
enableCompression: true # 启用记忆压缩v0.1.0 — Initial Release
@deepseek-ai/dsh-memory-connect v0.1.0
Cross-session memory plugin for DeepSeek Harness (DSH).
Features
- 🔍 Auto Extraction — Extracts facts, preferences, decisions, and context from conversations
- 🧠 Semantic Recall — RRF (Reciprocal Rank Fusion) combines keyword and semantic search
- ⏰ Scheduled Maintenance — Automatic periodic decay and consolidation
- 🤖 LLM Consolidation — Intelligent memory merging using DSH built-in
ctx.llm - 📉 Memory Decay — Old, unused memories naturally fade
- 🔧 Zero Config — Works with SQLite FTS5, no additional dependencies
Installation
npm install @deepseek-ai/dsh-memory-connectQuick Start
# agent.cordis.yml
- id: memory
name: @deepseek-ai/dsh-memory-connect
config:
path: ~/.dsh/memory.dbLinks
中文
新功能
- 🔍 自动提取 — 从对话中提取事实、偏好、决策和上下文
- 🧠 语义召回 — RRF 融合关键词和语义搜索
- ⏰ 定时维护 — 自动衰减和整合
- 🤖 LLM 整合 — 使用 DSH 内置 LLM 智能合并
- 📉 记忆衰减 — 旧记忆自然消退
- 🔧 零配置 — 基于 SQLite FTS5,开箱即用
安装
npm install @deepseek-ai/dsh-memory-connect