Adaptive Memory Graph

Load the lightweight memory graph index at session start. Returns a summary map of domains and top-weighted nodes — not the full graph. Use this at the start of every conversation for background awareness. Args: session_context: Optional domain filter (e.g. "health_and_safety") to focus the index on relevant domains.

Fetch the full content of a memory node when contextually relevant. Only call this when the conversation clearly warrants deeper context from a specific node. Updates the node's last_accessed timestamp. Args: node_id: The ID of the node to expand (from the index).

Get nodes connected to a given node above a minimum weight threshold. Useful for cross-domain discovery — finding related knowledge that might be relevant based on graph connections. Args: node_id: The node to find connections from. min_weight: Minimum connection weight to include (0.0-1.0).

Log session summary at the end of a conversation. Call this silently at the end of each conversation — do not narrate this to the user. Records which nodes were accessed, engagement signals, and any new node suggestions. Args: interface: "text" or "voice" branches_accessed: JSON array of node IDs that were accessed. branches_expanded: JSON array of node IDs that were fully expanded. engagement_signals: JSON array of objects with node_id, signal ("positive"/"neutral"/"negative"), and optional note. new_nodes_suggested: JSON array of objects with title, domain, summary, and optional connections_to array. explicit_corrections: JSON array of objects with node_id, correction text, and action ("decay"/"archive"/"update").

Process pending session logs and apply decay to the graph. This processes all pending session logs (engagement signals, corrections, new node suggestions) and applies time-based weight decay to all active nodes. Nodes that fall below 0.10 weight are archived. Typically called after each session or on a schedule.

Generate a human-readable summary of the current graph state. Shows active domains, top nodes, recent changes, and cross-domain connections. Useful for periodic review of how the memory graph is evolving. Args: format: Output format — currently only "markdown" is supported.

Manually adjust a node's weight or status. Allows the user to directly control their memory graph. Args: node_id: The node to adjust. action: One of "boost", "decay", "archive", "delete". value: Optional weight delta for boost/decay (e.g. 0.1 to boost by 0.1). Defaults to 0.1 for boost, -0.1 for decay.

Manually add a new node to the memory graph. Args: title: Short title for the node. domain: Domain category (e.g. "health_and_safety", "personal", "ideas_and_projects", "general"). summary: Brief summary used in the lightweight index. content: Full detailed content for the node. subdomain: Optional subcategory within the domain. tags: JSON array of string tags. file_links: JSON array of file path strings to link to this node. weight: Initial weight (0.0-1.0), defaults to 0.5. connect_to: JSON array of objects with node_id, weight, and label to create connections to existing nodes.

Search for nodes by title, summary, tags, or content. Args: query: Search text to match against node title, summary, content, and tags. domain: Optional domain filter. include_archived: Whether to include archived nodes in results.

List available Claude Code chat sessions for review. Scans ~/.claude/projects/ for conversation history. Returns session metadata including project name, date, message count, and a preview of the first user message. Use this to find sessions worth ingesting into the memory graph. Args: project_filter: Optional substring to filter by project name.

Read a Claude Code chat session and return its conversation content. Extracts user and assistant text messages from a session JSONL file, filtering out tool calls and system metadata. Use the output to identify knowledge worth adding to the memory graph via amg_add_node. Args: session_path: Full path to the session JSONL file (from amg_list_chat_sessions). max_messages: Maximum number of messages to return (default 100).