sessionmemory_skill

Overview

This skill prevents long-running agents from losing important details after context compaction by converting session logs into a three-layer, searchable memory system. It ships production-ready Python scripts to convert sessions to Markdown, auto-generate a structured glossary, and keep everything up-to-date via cron. The result is reliable navigation from high-level entities to full session transcripts so agents can find exact quotes, decisions, and timelines.

How this skill works

Scripts scan OpenClaw session logs and produce three memory layers: a curated MEMORY.md, an auto-generated SESSION-GLOSSAR.md (people, projects, topics, timeline, decisions), and per-session Markdown transcripts stored under memory/sessions/. The glossary indexes sessions and extracts entities, while optional cron jobs run the converter and glossary builder incrementally to maintain fresh, searchable context. Integrated search lets agents navigate from an entity to the precise session where details were discussed.

When to use it

• When agents forget specifics after automatic compaction or summarization.
• When you need reliable traceability of decisions, names, file paths, or dates.
• When running 24/7 agents, cron jobs, or many short-lived subagents that start with empty context.
• When multiple sessions, group chats, or cron jobs must share consistent project context.
• When you want an automated, maintainable memory layer rather than manual notes.

Best practices

• Run the two scripts once to populate memory, then schedule incremental cron jobs every 4–6 hours.
• Populate KNOWN_PEOPLE and KNOWN_PROJECTS in the glossary script to improve entity detection.
• Occasionally run a full glossary rebuild to pick up detection improvements and edge cases.
• Prepend a memory preamble to cron/subagent prompts to force a memory_search before work begins.
• Keep MEMORY.md curated for long-term facts and use the glossary for navigation to transcripts.

Example use cases

• A research cron job that needs recent context: run the cron-optimizer to add a memory preamble so it checks past findings before scraping.
• Spawning a subagent for code review: include memory_search keywords so the subagent finds prior decisions and style notes.
• Recovering an exact client instruction mentioned weeks ago by locating the session transcript via the glossary.
• Maintaining a project timeline and decision log across many short sessions, searchable by person or topic.
• Automating periodic checks that reference recent conversations without manual context injection.