memory-architectlisted

# memory-architect Designs an agent's memory along two axes: **what kind** of memory (type) and **how widely it is shared** (scope). The core discipline is deciding what lives *in context* (small, always present) versus *archival* (large, retrieved on demand) — because context is the scarce resource. ## When to use - An agent forgets across turns/sessions, or its context is bloated with stale facts. - Deciding what to keep in the prompt vs. retrieve on demand. - Choosing whether a fact is per-conversation or shared across all of a user's sessions. ## Run it ``` python -m harness.memory demo python .claude/skills/memory-architect/scripts/run.py demo ``` The demo walks working memory → promote to archival → vector recall, all in-memory and offline. ## Two axes (detail in `reference.md`) - **Type**: `working` (in-context scratchpad), `factual` (stable facts), `episodic` (events that happened), `semantic` (distilled knowledge). - **Scope**: `conversation` → `session` → `user` → `org` (narrowest to widest sharing). ## In-context vs archival - **Core / in-context**: tiny, always in the prompt (current task, key facts). `core_append`, `core_view`. - **Archival**: large, stored out-of-context, retrieved by similarity. `archival_insert`, `archival_search`. - **promote**: move a working item across the boundary into durable archival memory when it proves worth keeping. ## Method 1. Default everything to working/conversation scope; promote only what recurs. 2. Keep the in-co