菜单
Automatically loads recent conversation memory into new sessions and generates AI summaries during compaction to maintain continuity across conversations. Preserves the last 10 raw messages verbatim so the agent can resume exactly mid-conversation without paraphrasing loss.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/knownasnaffy/prompthound --skill session-memory-summarization命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Install and operate local NVIDIA Parakeet ASR for OpenClaw with an OpenAI-compatible transcription API on Ubuntu/Linux and macOS (Intel/Apple Silicon). Use when the user wants private/local speech-to-text, voice transcription setup, ASR troubleshooting, or OpenClaw voice stack configuration with Parakeet (and optional Whisper fallback).
Collect, refresh, normalize, and analyze the user's own Douban history for taste analysis and recommendation reasoning. Use when the task involves the user's own Douban shelves, ratings, tags, comments, reviews, or recent activity, especially when you need to decide whether local cache is fresh enough, re-crawl logged-in data with cookies, store refreshed results locally, and then analyze them by category.
LoRA fine-tuning pipeline for Stable Diffusion on Apple Silicon — dataset prep, training, evaluation with LLM-as-judge scoring. Use when fine-tuning image generation models for consistent style, custom characters, or domain-specific visuals. Requires Python with torch and diffusers.
基于 SOC 职业分类
正在显示 SKILL.md
| name | Session Memory & Summarization |
| slug | session-context |
| description | Automatically loads recent conversation memory into new sessions and generates AI summaries during compaction to maintain continuity across conversations. Preserves the last 10 raw messages verbatim so the agent can resume exactly mid-conversation without paraphrasing loss. |
| author | AniBot (Thomas) |
| version | 0.2.0 |
| license | MIT |
| tags | ["memory","session","summarization","continuity","ai"] |
| hooks | ["session:start","session:compact:before"] |
| repository | https://github.com/animo66/openclaw-skills |
| homepage | https://clawhub.ai/skills/session-context |
| minOpenClawVersion | 0.29.0 |
python scripts/_secrets_inline.py is review material for the Session Memory & Summarization Skill skill. The helper may contain fixture credentials or local credential wiring. Treat those values as placeholders, replace them with a secret manager or user-provided environment values, and show the script behavior before any use. Do not contact external services with embedded sample credentials.
Provides automatic conversation continuity across sessions by loading recent memory at session start and generating AI summaries during compaction.
session:startRuns when a new session begins. Loads the most recent daily memory file and injects two context blocks:
session:compact:beforeRuns before automatic compaction (20+ messages OR 60% of token limit). Does two things:
agent.generateSummary() and prepends it to today's memory file<!-- recent_messages_block -->) — this is what session:start reads back next sessionclawhub install session-context
Or manually:
cd ~/.openclaw/workspace/skills
git clone https://github.com/thomasmarcel/openclaw-skill-session-context.git session-context
openclaw skills enable session-context
memory/ directory (created automatically)Customize thresholds in hooks/session/compact:before/handler.js:
return (
msgCount >= 20 || // minimum messages before summarizing
tokenCount > maxTokens * 0.6 // trigger at 60% of token limit
);
Adjust how many raw messages to preserve:
// In both handler files:
const MAX_RECENT_MESSAGES = 10; // last N user/assistant turns to preserve verbatim
Adjust summary context size:
// In hooks/session/start/handler.js:
const MAX_SUMMARY_CHARS = 6000; // cap on AI summary injected at session start
memory/
2026-04-03.md # daily files — summaries at top, recent_messages block at bottom
2026-04-04.md
Each file has this structure:
## HH:MM:SS
<AI summary of the session>
---
## Earlier timestamp
<earlier summary>
<!-- recent_messages_block -->
[{"role":"user","content":"..."},
{"role":"assistant","content":"..."},
...]
The <!-- recent_messages_block --> section is always at the end and replaced each compaction with the latest N turns.
session:compact:before hook checks thresholds. If met:
memory/YYYY-MM-DD.mdsession:start hook loads the file and injects both:
MIT