reme-remelight

name	reme-remelight
description	Use when integrating ReMeLight into an agent loop and you need file-based context checks, compaction, tool-result offload, semantic search, or session memory.

ReMe ReMeLight

Use this skill for the file-based runtime capabilities exposed by ReMeLight.

When to Use

Use this skill when you need to:

• keep long conversations within model context limits
• compact history into structured summaries
• offload oversized tool results into files
• run semantic search over file-based memory
• create session-scoped in-memory state that still persists dialog history

Do not use this skill for vector-backed personal/task/tool memory flows or for HTTP and MCP startup concerns.

Core Surface

The main runtime methods are:

• start / close
• check_context
• compact_memory
• compact_tool_result
• pre_reasoning_hook
• summary_memory
• memory_search
• get_in_memory_memory
• await_summary_tasks

Working Model

Treat ReMeLight as a file-based runtime manager:

• memory/ stores summarized durable notes
• dialog/ stores raw conversation records
• tool_result/ stores offloaded large tool outputs

Recommended order in an agent loop:

1. start the runtime
2. compact large tool outputs before reasoning
3. check token pressure
4. compact history when thresholds are exceeded
5. trigger summary_memory for durable facts
6. wait for background summary tasks before shutdown when required

Design Rules

• keep file-based runtime concerns separate from vector-memory concerns
• use pre_reasoning_hook when you want the standard orchestration path
• use individual methods only when you need tighter control over thresholds or sequencing
• do not bind this skill to a private workspace path