Agentcore MCP — VS Code Extension

Persistent local memory for AI agents in VS Code.

Agentcore gives Copilot, Cursor, and MCP-compatible assistants a local memory layer for tasks, events, decisions, lessons, facts, search, and workspace handoff. It also gives you a VS Code dashboard so memory is visible to the human, not only callable by the AI.

No cloud account required for the default local path
No API key required for local memory
Single-user local storage at ~/.agentcore/agentcore.sqlite by default
Optional Azure AI Search can be configured later for deep_search

Quick start

Install Agentcore MCP from the VS Code Marketplace.
Open the VS Code Command Palette:
- macOS: Cmd+Shift+P
- Windows/Linux: Ctrl+Shift+P
- If you use Quick Open (Cmd+P / Ctrl+P), type > first, then search Agentcore.
Run Agentcore: Show Welcome.
Run Agentcore: Show Health to confirm the backend and database path.
Run Agentcore: Install AI Memory Rules so Copilot/Cursor know when to read and write Agentcore memory.
Open Copilot Chat in Agent mode and keep working normally.

If the dashboard looks empty at first, that is expected until an AI assistant or a capture command records tasks, decisions, lessons, questions, or events.

Main UI commands

Open these from the Command Palette by searching Agentcore.

Command	What it does
`Agentcore: Show Welcome`	Opens the onboarding page.
`Agentcore: Show Health`	Shows active backend, database path, authority/lock status, backups, and startup diagnostics.
`Agentcore: Show Memory Dashboard`	Opens the read-only memory board: active tasks, completed tasks, open questions, decisions, lessons, recent activity, keyword search, and storage paths.
`Agentcore: Resume Workspace`	Builds a concise workspace-memory snapshot and includes a Copy as Prompt action for chat handoff.
`Agentcore: Install AI Memory Rules`	Adds managed Copilot/Cursor instruction blocks so the assistant reliably uses Agentcore tools.
`Agentcore: Save Decision`	Manually records a durable decision.
`Agentcore: Save Lesson`	Manually records a reusable lesson.
`Agentcore: Mark Task Complete`	Marks an active task complete from VS Code.
`Agentcore: Clean Demo/Test Memory`	Safely removes demo/test memory after creating archive/backup protection.
`Agentcore: Export JSON Snapshot`	Exports a portable JSON snapshot of local memory.
`Agentcore: Open Data Directory`	Opens the local Agentcore data folder.
`Agentcore: Open Backup Directory`	Opens the backup folder.
`Agentcore: Copy Database Path`	Copies the active database path.
`Agentcore: Open Operational Log`	Opens the local operational log.
`Agentcore: Show Backend Diagnostics`	Shows detailed backend and startup diagnostics.

How the memory board works

Agentcore: Show Memory Dashboard is the human-facing board for local memory. It is read-only and filters seed/test rows from the default view.

It shows:

active tasks
recently completed tasks
open questions
recent decisions
recent lessons
recent activity
keyword search over event content
backend, database, backup, and authority status

The board is populated by MCP tools such as start_task, update_task, complete_task, ask_team, save_decision, save_lesson, log_event, and submit_feedback. It can also be populated by the manual VS Code capture commands listed above.

How AI assistants should use Agentcore

Use normal workspace/source-code tools for files and code. Use Agentcore for memory that should survive across chat reloads and sessions.

Recommended behavior:

At session start or when the user asks “where are we?”, call session_snapshot.
For recall, call get_facts first, then search with concrete phrases, IDs, or error text.
Use deep_search only after local search misses or after Azure AI Search has been configured.
During execution, record meaningful progress with task lifecycle tools.
Save durable decisions and lessons with save_decision and save_lesson.
Use submit_feedback when Agentcore UX, docs, or tool behavior creates friction.
If scope is ambiguous after one sensible recall pass, ask one short clarifying question instead of guessing.

MCP tool layers

Layer	Tools
L0 Orient	`session_snapshot`, `get_tasks`, `get_questions`, `search`, `deep_search`, `get_agentcore_routing`
L1 Execute	`log_event`, `start_task`, `update_task`, `complete_task`, `fail_task`, `block_task`, `rework_task`
L2 Structure	`add_entity`, `get_facts`, `get_linked`, `save_decision`, `save_lesson`, `warm_start`
L3 Collaboration / portability / feedback	`ask_team`, `answer_question`, `resolve_question`, `export_data`, `import_data`, `submit_feedback`

Privacy and storage

By default, Agentcore stores memory locally on your machine:

Data directory: ~/.agentcore
Database file: agentcore.sqlite
Default backend: auto, which prefers native SQLite when available and falls back to local sql.js snapshot mode

Agentcore does not require a cloud account or API key for local memory. Optional Azure AI Search settings are only used if you configure them.

Troubleshooting

I cannot find Agentcore commands

Make sure you are using the VS Code Command Palette:

Cmd+Shift+P on macOS
Ctrl+Shift+P on Windows/Linux
Or Cmd+P / Ctrl+P, then type > before searching Agentcore

If commands still do not appear, run Developer: Reload Window.

Dashboard opens but is empty

This usually means no user memory has been recorded yet. Run Agentcore: Install AI Memory Rules, then work with Copilot/Cursor in Agent mode so the assistant knows to call Agentcore tools. You can also add memory manually with Agentcore: Save Decision, Agentcore: Save Lesson, or Agentcore: Mark Task Complete.

MCP fails to start

Run Agentcore: Show Health. If the backend is misconfigured, set agentcore.storageBackend to auto or local. The cosmos backend is reserved for a future release and is not implemented yet.

I want to inspect or back up my data

Use:

Agentcore: Open Data Directory
Agentcore: Open Backup Directory
Agentcore: Copy Database Path
Agentcore: Export JSON Snapshot

Key settings

Setting	Default	Purpose
`agentcore.storageBackend`	`auto`	Prefer native SQLite when supported; fall back to local snapshot mode.
`agentcore.dataDir`	`~/.agentcore`	Local data directory.
`agentcore.dbFileName`	`agentcore.sqlite`	SQLite database filename.
`agentcore.defaultAgentId`	`copilot-agent`	Default agent ID used when logging events from VS Code commands.
`agentcore.searchDefaultLimit`	`10`	Default result count for search-style tools.
`agentcore.taskDefaultLimit`	`20`	Default task count for task-board queries.
`agentcore.warmStartBudgetDefault`	`8000`	Default token budget for `warm_start`.
`agentcore.azureSearchEndpoint`	(empty)	Optional Azure AI Search endpoint for `deep_search`.
`agentcore.azureSearchKey`	(empty)	Optional Azure AI Search key.
`agentcore.azureSearchIndex`	`bronze-events`	Azure AI Search index name.
`agentcore.azureSearchSemanticConfiguration`	`default`	Clear this setting to use simple Azure keyword search only.

Current boundary

Agentcore is currently a local-first, single-user desktop memory layer. It is not yet a shared team memory service.

License

MIT

Agentcore MCP

David Jinguo Xu