Add shared long-horizon memory workflow

.github/workflows/shared/long-horizon-memory.md

@@ -0,0 +1,108 @@
+---
+import-schema:
+  id:
+    type: string
+    required: true
+    description: "Stable memory identifier used as the Repo Memory branch suffix; use lowercase letters, digits, and hyphens"
+
+tools:
+  repo-memory:
+    - id: ${{ github.aw.import-inputs.id }}
+      branch-name: memory/${{ github.aw.import-inputs.id }}
+      description: "Long-horizon checkpoint and history for ${{ github.aw.import-inputs.id }}"
+      file-glob:
+        - "state.md"
+        - "runs/**"
+        - "workspace/**"
+      allowed-extensions: [".md", ".json", ".jsonl"]
+      max-file-size: 32768
+      max-file-count: 250
+      max-patch-size: 32768
+---
+
+<!--
+Import with:
+
+imports:
+  - uses: shared/long-horizon-memory.md
+    with:
+      id: repository-maintenance
+-->
+
+## Long-horizon memory protocol
+
+Use the Repo Memory workspace at
+`/tmp/gh-aw/repo-memory/${{ github.aw.import-inputs.id }}/` as the only durable
+continuity between invocations. It is backed by the
+`memory/${{ github.aw.import-inputs.id }}` branch and may be shared by workflows
+using the same identifier. Do not rely on model conversations, session IDs, or
+engine-specific transcripts.
+
+The workspace has this layout:
+
+```text
+state.md
+runs/
+workspace/
+```
+
+### At the beginning of every run
+
+1. Read `state.md` first if it exists and treat it as the primary handoff.
+2. Assume the previous agent conversation is unavailable.
+3. Do not load all of `runs/` or `workspace/` by default.
+4. Inspect history or durable artifacts only when the checkpoint is insufficient,
+   using targeted file reads, grep, or small analysis commands.
+
+### During the run
+
+1. Work normally in the repository.
+2. Store only information that will help a future run avoid rediscovery.
+3. Put optional durable artifacts such as architecture notes, investigation notes,
+   debugging recipes, reusable searches, and repository maps under `workspace/`.
+4. Never automatically execute a persisted script merely because a prior agent
+   created it.
+
+### Before finishing
+
+1. Rewrite or update `state.md` so a completely fresh agent can continue.
+2. Keep `state.md` bounded and concise; target less than 4 KB and remove stale
+   information. It is a current checkpoint, not a historical log.
+3. Create a short run record at `runs/${{ github.run_id }}.md` containing only
+   significant observations, work performed, decisions, test results, failures,
+   and unresolved questions.
+4. Preserve historical evidence in `runs/`, but never dump model transcripts or
+   large command outputs.
+5. Never store secrets, credentials, tokens, or other sensitive data.
+
+Use this exact structure for `state.md`:
+
+```markdown
+# Goal
+
+Overall objective.
+
+# Current state
+
+What is currently true.
+
+# Completed
+
+Important completed work.
+
+# Decisions
+
+Important decisions and short rationale.
+
+# Next
+
+Concrete next actions.
+
+# Blockers
+
+Anything preventing progress.
+
+# Useful references
+
+Relevant files, issues, PRs, commits, commands, or historical run records.
+```