diff --git a/marketing/2026-07-13-cross-host-local-learning-brief.md b/marketing/2026-07-13-cross-host-local-learning-brief.md
new file mode 100644
index 0000000..a8b9bde
--- /dev/null
+++ b/marketing/2026-07-13-cross-host-local-learning-brief.md
@@ -0,0 +1,86 @@
+# claude-smart marketing brief — cross-host local learning (2026-07-13)
+
+## Why this run
+
+`origin/main` now clearly presents claude-smart as a learning layer for Claude Code, Codex, and OpenCode. The current public conversation around coding-agent memory is crowded with "shared memory", "second brain", and "MCP memory" positioning. The strongest low-risk angle for claude-smart is narrower and more differentiated:
+
+> claude-smart turns corrections and successful workflows into local Preferences, Project-specific skills, and Shared skills that Claude Code, Codex, and OpenCode can reuse in future sessions.
+
+This avoids generic memory claims while highlighting the parts developers can verify from the repo: local install, dashboard/auditability, host support, and rule-like learned artifacts.
+
+## Read-only market signals checked
+
+- GitHub repo metadata is aligned: description mentions Preferences, Project-specific skills, Shared skills, and topics include `claude-code-plugin`, `codex-plugin`, `opencode-plugin`, `agent-memory`, `ai-agent-memory`, `local-first`, and `self-improving-ai`.
+- X/Twitter search was used read-only only. No likes/replies/posts/DMs were sent.
+- Recent X results for `"Claude Code" memory`, `"Claude Code" plugin memory`, and `"Codex" memory agent` show active discussion around:
+  - Claude Code agents forgetting between sessions.
+  - Local/shared memory across Claude Code, Codex, and other coding agents.
+  - Obsidian/second-brain memory workflows.
+  - MCP-backed memory tools.
+
+## Positioning recommendation
+
+Lead with **learning from corrections**, not "another memory store".
+
+Suggested hierarchy:
+
+1. **Problem:** coding agents repeat the same mistakes across sessions and hosts.
+2. **Mechanism:** claude-smart distills corrections and successful workflows into Preferences, Project-specific skills, and Shared skills.
+3. **Distribution:** one local learning layer works across Claude Code, Codex, and OpenCode.
+4. **Proof hooks:** inspectable dashboard, local storage, and benchmark in `EXPERIMENT.md` for correction capture vs `claude-mem`.
+
+## Ready-to-post first-party update
+
+Use after any active release/install issue is resolved, especially if issue #130 is still open when posting.
+
+```text
+Coding-agent memory should do more than remember a transcript.
+
+The useful loop is:
+correction → Project-specific skill → Shared skill → future Claude Code/Codex/OpenCode sessions start from the better path.
+
+That is the design behind claude-smart: a local learning layer powered by Reflexio.
+
+Repo: https://github.com/ReflexioAI/claude-smart
+```
+
+## Suggested X post for Yi to review
+
+No X action was taken automatically.
+
+```text
+A pattern I keep seeing in coding-agent memory tools: they optimize for storing more context.
+
+For coding, the higher-leverage unit is often a future-facing rule:
+"when this situation appears again, do X instead of Y."
+
+claude-smart turns corrections and successful workflows into Preferences, Project-specific skills, and Shared skills that Claude Code, Codex, and OpenCode can reuse locally.
+
+https://github.com/ReflexioAI/claude-smart
+```
+
+## Suggested X reply for relevant threads
+
+Only use when someone is explicitly asking about Claude Code/Codex/OpenCode memory or repeated agent mistakes. Do not use as a cold promo reply.
+
+```text
+One framing that helped us: store less transcript, more future-facing rule.
+
+claude-smart turns corrections into Preferences, Project-specific skills, and Shared skills, so future Claude Code/Codex/OpenCode sessions can start from the learned path instead of rediscovering it.
+
+Repo: https://github.com/ReflexioAI/claude-smart
+```
+
+## Non-X distribution opportunities
+
+1. **GitHub issue #55 follow-up:** ask users to share one stale or useful Project-specific skill after the latest host-support changes. Keep it framed as product feedback, not marketing.
+2. **Docs/blog outline:** "Memory vs learning for coding agents" — use examples already in `README.md` and tie benchmark claims only to `EXPERIMENT.md`.
+3. **OpenCode communities:** if Yi has an owned/approved channel, share the one-sentence install update: `npx claude-smart install --host opencode` plus the note that local Preferences, Project-specific skills, and Shared skills are reused across supported hosts.
+
+## Guardrails
+
+- Do not claim hosted sync, team sharing, customer adoption, or benchmark results beyond `EXPERIMENT.md`.
+- Do not post on X/Twitter automatically.
+- Avoid competitor attacks; use the factual "memory vs learning" distinction.
+- Preserve public terminology: Preference, Project-specific skill, Shared skill.
+- If release/install issue #130 is still open, avoid broad launch amplification and keep messaging to docs/feedback channels until fixed.