Memory House - Agent Plugin (POC)

Packages the deployed Memory House service as an Agent Plugin 1.0 for the GitHub Copilot app (also VS Code and the CLI). It bundles:

• MCP server (mcp.json) - the memory-house server with its 12 tools, via the IP gateway (OAuth sign-in handled by the client, no token in the file).
• Skills (skills/use-memory/, skills/mh-login/) - teach the agent to use memory and provide the /mh-login workflow.
• Hooks (hooks/hooks.json) - deterministic recall and capture that MCP alone cannot do:
  • userPromptSubmitted -> inject.sh captures the sanitized user turn.
  • userPromptTransformed -> inject.sh retrieves memories and adds them to the model-facing prompt.
  • postToolUse on enroll_hook_capture -> redeems the credential locally and replaces the tool result so the credential cannot appear in the agent response.
  • agentStop -> capture.sh appends the turn to Memory House.
• Commands (com.github.copilot/commands/) - /mh-show, /forget.
• Canvas (com.github.copilot/extensions/memory-canvas/) - a "Memory" panel that shows what Memory House remembers about you, grouped by Personal / Team / Org, with refresh / forget / promote actions. See its own README.

The backend (gateway + managed service + core + durable pipeline + Cosmos) is unchanged; this is purely a client-surface package.

Layout

plugin/
├── plugin.json
├── mcp.json
├── hooks/hooks.json
├── skills/{use-memory,mh-login}/SKILL.md
└── com.github.copilot/
    ├── scripts/{inject,capture,complete-login,amt-token}.sh
    ├── commands/{mh-show,forget}.md
    └── extensions/memory-canvas/{package.json,extension.mjs,README.md}

Prerequisites

• jq and curl on PATH (used by the macOS/Linux hook scripts).
• Run /mh-login once after connecting the MCP server so the hooks have a gateway-issued access/refresh token.
• The scripts are executable: chmod +x com.github.copilot/scripts/*.sh.

What is demo-grade vs. real

• Auth (mh-token.sh): reads and silently refreshes the hook token enrolled by /mh-login; it does not depend on an interactive Azure CLI session.
• Login completion: /mh-login only calls enroll_hook_capture; complete-login.sh performs redemption deterministically in postToolUse. It does not depend on the model remembering a second step or on opening the canvas first.
• Hook lifecycle: userPromptSubmitted performs capture as a side effect and returns {}; current config-file hooks discard its output. Recall runs in userPromptTransformed and returns modifiedTransformedPrompt, as defined by the Copilot hooks reference.
• Diagnostics: hook invocations and non-sensitive outcomes are appended to ~/.copilot/amt/hook.log; prompts, memories, and tokens are never logged there.
• Windows: PowerShell hook entries and .ps1 twins are included.
• /forget: no delete endpoint exists yet (Memory House supersedes, not deletes).

Quick local check (no Copilot needed)

Verify the scripts talk to Memory House with your identity:

# user capture phase
echo '{"sessionId":"plugin-smoke","prompt":"test user turn from the plugin"}' \
  | AMT_HOOK_PHASE=capture ./com.github.copilot/scripts/inject.sh

# recall phase: what would be sent to the model
echo '{"sessionId":"plugin-smoke","prompt":"what did we decide about signing x-amt-context?","transformedPrompt":"what did we decide about signing x-amt-context?"}' \
  | AMT_HOOK_PHASE=recall ./com.github.copilot/scripts/inject.sh

# capture path: append the last agent message from a Copilot-style transcript
printf '%s\n' '{"type":"assistant.message","data":{"content":"test turn from the plugin"}}' \
  > /tmp/amt-plugin-smoke.jsonl
echo '{"sessionId":"plugin-smoke","transcriptPath":"/tmp/amt-plugin-smoke.jsonl"}' \
  | ./com.github.copilot/scripts/capture.sh
rm /tmp/amt-plugin-smoke.jsonl

Install (GitHub Copilot app / CLI / VS Code)

MCP-only path works today with just mcp.json (add the server URL in the app's Customize tab; sign in when prompted). Full plugin install (with hooks + commands) follows the Agent Plugins install flow for your client; point it at this plugin/ directory. Hooks run locally, so jq and curl must be available in the shell the client uses.

See Docs/amt-plugin-design-sketch.md for the full design, the canvas ("see my memories") surface, and the effort breakdown.