memory-system-sidecar（内存系统边车）

Memory System Sidecar

Use this skill for the implemented MemoryLab memory system, not just abstract planning.

The system boundary is:

• Context manager = hot-path compactor
• Memory sidecar = long-term extraction, indexing, retrieval, consolidation
• Bridge = feeds retrieved memory back into ACTIVE_TASK_STATE / LIVE_CONTEXT_PACKET without treating them as source of truth

Quick workflow

Choose the smallest workflow that matches the task.

1. Refresh live context + sidecar from a real session

Use when the user says things like:

• "refresh memory"
• "rebuild context from the laptop session"
• "update the live packet from the current session"

Run:

./skills/memory-system-sidecar/scripts/refresh_memory_system.sh

This wraps the repo’s standard end-to-end refresh path.

2. Verify the memory system still passes

Use when the user asks:

• "is the memory system working?"
• "does milestone 5 still hold?"
• "run the checks"

Run:

./skills/memory-system-sidecar/scripts/verify_memory_system.sh

This runs the main unit tests plus the eval harness.

3. Rebuild indexes only

Use when eval/retrieval drifts or docs mention stale-index behavior.

Run:

./skills/memory-system-sidecar/scripts/rebuild_memory_indexes.sh

Then rerun verify.

Operating rules

• Treat raw history / transcripts as source of truth.
• Treat memory/LIVE_CONTEXT_PACKET.md as working memory, not durable storage.
• Do not treat retrieved-memory injection as new durable truth.
• Prefer repo wrappers (scripts/context-refresh*) over ad-hoc manual command chains.
• When status docs and runtime disagree, trust runtime checks after rebuild.

Read these references when needed

• references/commands.md — exact command map for refresh / rebuild / verify / inspect
• references/docs-map.md — which doc is design vs status vs restore vs checklist

Common files to inspect

• memory-system/MILESTONE_STATUS.md
• docs/designs/MEMORYLAB_MEMORY_SYSTEM_MILESTONE_RESTORE.md
• docs/designs/MEMORYLAB_MEMORY_SYSTEM_BLUEPRINT.md
• memory/ACTIVE_TASK_STATE.md
• memory/LIVE_CONTEXT_PACKET.md
• memory/RETRIEVAL_CONTEXT_BUNDLE.json
• history/transcripts/*.md

Default debug order

When output looks wrong, inspect in this order:

1. memory/ACTIVE_TASK_STATE.md
2. memory/LIVE_CONTEXT_PACKET.md
3. memory/RETRIEVAL_CONTEXT_BUNDLE.json
4. history/transcripts/*.md
5. memory-system/index/memory.db freshness / rebuild state

Compatibility rule

This skill operates the real sidecar already implemented in this repo. It should complement:

• context-manager
• live-context-management
• long-term-memory

Use this skill when the task is about running or auditing the built system, not just designing one.