memory-admin

Installation
SKILL.md

Purpose

Use this skill only for maintenance and diagnostics of the unified project memory system.

Typical tasks:

  • run catalog repair
  • inspect current code changes for memory-worthy decision candidates
  • inspect manifest / store layout
  • diagnose index drift or missing refs
  • run transcript distillation through the CLI extractor

Scope

This skill is not the primary workflow for reading or writing memory.

Primary workflow:

  • project-memory

Allowed Actions

  • memory-hub catalog-repair
  • memory-hub discover [--summary-file <summary>]
  • memory-hub session-extract --file <session-transcript>
  • inspect .memory/manifest.json
  • inspect .memory/docs/, .memory/catalog/, .memory/_store/
  • inspect .memory/_store/projections/boot.json and search.json
  • diagnose CLI / MCP behavior issues

When this skill needs repository memory context, use the memory-hub MCP tools directly:

  • read_memory(ref="catalog://topics")
  • read_memory(ref="doc://...")
  • search_memory(query, scope=docs|all, ...)

Do not treat catalog://... or doc://... as MCP resources. Do not call read_mcp_resource, and do not invent a server name such as memory.

Session Distillation

Use memory-hub session-extract --file <session-transcript> when the user asks to distill a completed conversation into memory candidates.

This is a CLI maintenance action, not an MCP tool call. Do not invent session_extract(...) or any other MCP tool name for it.

The extractor:

  • classifies transcript chunks into docs-only / durable-only / dual-write
  • routes them through the existing unified write lane
  • creates docs change reviews or durable proposals instead of writing active state directly

Do not use this skill to bypass the normal write lane.

Decision Discovery

Use memory-hub discover when the user asks whether the current code changes contain new rules, exceptions, or docs drift worth turning into memory candidates.

This command is read-only. It does not write active docs or approved durable memory. It only returns structured candidates with:

  • why the candidate is worth capturing
  • which changed files and existing refs support it
  • whether it likely belongs in docs-only, durable-only, or dual-write

If the user wants to continue after reviewing candidates, route the chosen item through the existing unified write lane instead of inventing a direct apply path.

Related skills
Installs
1
First Seen
Apr 14, 2026