docs+test: individual-agent positioning, Cursor as a host, and a real MCP smoke

[kevintseng]

Summary

Two commits that were made directly on a local main and had never been through CI. Moved onto a branch and opened as a PR without altering their content — main was reset back to origin/main, so nothing was lost and nothing was pushed unreviewed.

1. docs: clarify local cross-agent memory scope — repositions the product from "Claude Code & MCP coding agents" plus team sharing, to agentic memory for one developer's AI coding agents, and documents Cursor as a first-class MCP host.
2. test: add packaged MCP protocol smoke — the packaged-artifact smoke test now completes a real MCP exchange instead of stopping at "the database opens".

Type of change

• Feature (feat)
• Bug fix (fix)
• Refactor (refactor)
• Docs (docs) — commit 1
• Test (test) — commit 2
• Build / CI / chore
• Release (release)

What changed

Positioning (README ×3, docs/ARCHITECTURE.md, docs/api/API_REFERENCE.md)

before	after
agentic memory for Claude Code & MCP coding agents	agentic memory for individual AI coding agents, compatible with Claude Code, Codex, Gemini, Cursor
📦 Team Sharing — export → share with your team	📦 Personal backup and migration — export → another machine
export: "Share memories as JSON between projects or team members"	"Back up, migrate, or move memories as JSON between compatible agents"
### Team Sharing Workflow, --namespace team	### Personal Backup and Cross-Agent Transfer, --namespace personal
testimonial: "3-person startup, shared knowledge base"	testimonial: solo developer recalling from Codex what Claude Code stored

Cursor as a documented host — ~/.cursor/mcp.json (global) or .cursor/mcp.json (project) in README ×3, a new section 5 in llms-install.md with its own failure/remedy table, and a row in the ARCHITECTURE integration-tier table.

Packaged MCP protocol smoke (scripts/smoke-packed-artifact.mjs) — the test installed the tarball outside the repo and stopped at openDatabase. It now speaks the wire protocol from the installed consumer tree: stdio initialize, listTools asserted to be exactly the nine shipped tools, then remember followed by recall with the written memory asserted present in the response. Importing the handlers directly would have missed the handshake, the tool-list schema and the server lifecycle every host depends on — which is the gap this closes.

Docs synced

• CHANGELOG.md — not updated. Positioning and doc coverage, not shipped behaviour; the smoke test is CI-side. Flagging it rather than deciding silently: if the Cursor support should read as a user-facing addition, it wants an entry.
• docs/ARCHITECTURE.md updated — Cursor row in the integration table, and the team-sharing workflow section rewritten
• docs/api/API_REFERENCE.md updated — export description
• README.md updated
• README locales re-synced — README.zh-TW.md and README.de.md carry the same edits
• Version files — no bump
• dist/skills-manifest.json — nothing under .claude-plugin/, scripts/hooks/, skills/ or version files was touched
• memesh doctor — not run

Verification

I did not run these locally and am not claiming they pass. These two commits were authored in another session; my part was moving them off main onto a branch, reading both diffs in full, and opening this PR. CI is what verifies them, and two of its steps bear directly on this content:

• Release Verification Gate → check-doc-claims.mjs, which validates every CLI flag, subcommand and tool name the public docs mention. --namespace personal and the nine tool names in the new smoke test are exactly its subject matter.

• Packaged Artifact Smoke Test → runs the new MCP exchange against a real packed tarball.

• npx tsc --noEmit — not run locally

• npm run build — not run locally; no source under src/ changed

• npm test -- --run — not run locally

• Hooks untouched

• LLM flows untouched

Test plan

• node scripts/smoke-packed-artifact.mjs — packs, installs, and completes initialize → listTools → remember → recall
• node scripts/check-doc-claims.mjs — every flag and tool name in the changed docs resolves
• Register memesh-mcp in Cursor per the new README section and confirm the tools appear

Known limitations / follow-ups

• The team namespace still exists in the code (personal | team | global) while the docs now describe personal use. Not a contradiction — the namespace still works — but the two will drift if the positioning holds and nobody revisits it.
• The Cursor instructions are documentation of an external product's config format; nothing in this repository tests that they are current.