Aide Memory

What aide-memory MCP server does

The aide-memory MCP server gives coding agents a persistent memory layer alongside existing rules files. It captures knowledge that would otherwise remain in a conversation, including contributor preferences, technical facts, decisions tied to a code area, and team guidelines. Memories can be scoped with path globs such as src/components/dashboard/**, allowing the agent to receive context related to the files it is working on rather than loading one large rules file every time.

The project stores memory content in .aide/memories/ as JSON files. Shared memories can be committed to a repository and distributed through Git, while personal preferences can be kept in preferences/personal/, which is gitignored by default. Claude Code and Cursor use the same memory directory, so knowledge captured in one can be available in the other.

How it works

Running initialization installs six lifecycle hooks. A prompt-submission hook looks for corrections and can prompt the agent to save them. A stop hook periodically asks whether decisions or findings should be retained. At session start, selected preferences, guidelines, and priority memories are injected. Before file reads or edits, the system checks matching scopes and prompts recall when relevant memories exist. Post-tool tracking prevents repeated prompts for the same recalled information, while pre-compaction tracking resets that state after context compaction.

Recall ranking favors area_context, followed by technical, preferences, and guidelines. Scoped memories rank above project-wide entries. The aide-memory MCP server also supports keyword search through FTS5, with optional semantic search using Transformers.js or Ollama.

Setup and configuration

The documented installation is:

npm install -g aide-memory
aide-memory init

The project requires Node.js 18 or later and npm or npx. Initialization creates the project configuration and wires the hooks into supported editor workflows. After initialization, restart the editor or start a fresh session so the MCP server registers. Cursor users must also enable aide-memory in the editor's MCP settings.

Configuration is seeded in .aide/config.json. Available settings control reflection frequency, scope thresholds, session-start context size, active hooks, correction detection, and whether preferences are personal or shared. Telemetry can be disabled with AIDE_TELEMETRY=off.

Tools and capabilities

The MCP interface exposes seven tools:

• aide_recall retrieves relevant memories for a path or context.
• aide_remember stores a new memory.
• aide_update changes an existing memory.
• aide_forget removes a memory.
• aide_search searches stored memories.
• aide_memories lists available memories.
• aide_import imports memory data.

The CLI provides corresponding memory operations plus initialization, statistics, recall-log inspection, configuration, synchronization, migration, and cleanup commands.

Limitations and notes

Memory content remains on the local machine, but anonymized usage data is sent to PostHog by default. The reported data includes event type, a hashed machine identifier, platform, and Node version; telemetry can be turned off. The project does not require Docker, an external database, an API key, or a cloud account.

Claude Code is the reference adapter, and Cursor has full hook and MCP wiring. Windsurf, Codex, and Copilot receive a rules template at launch, but deeper integration is described as dependent on future user feedback. The project is free to use, but the provided material does not name a specific software license.