Genesys

Record a correction or update to an existing memory. Creates a NEW memory containing the corrected content and links it to the old one with a SUPERSEDES edge — the old memory is kept for provenance but automatically ranked down and tagged superseded_by in future recalls. Use when a fact changes ('my rate is now $150', 'we moved the deadline'), instead of storing a disconnected duplicate. Only the memory's owner can amend it.

memory_amend

Permanently delete a memory node and all its edges. Call when the user says 'forget this', 'delete that memory', 'remove that', or wants specific information erased. Always confirm with the user before deleting.

delete_memory

Explain a memory's score breakdown, causal basis, and removal impact.

memory_explain

List all pinned/core memories, optionally filtered by category, sorted by causal weight. Call when the user asks 'what do you know about me?', 'show me my memories', 'what have you saved?', or wants an overview of stored information.

list_core_memories

Get graph statistics: node counts by status, edge counts by type, orphan count.

memory_stats

Pin a memory to core status, preventing it from being forgotten.

pin_memory

Recall memories using hybrid search (vector + keyword + graph spreading activation). Call this when the user references stored context or past conversations ('do you remember', 'what did I say about'), or asks a question their stored history, preferences, or past decisions would clearly improve. Also call before memory_store to find related memories to link. Results are ranked by decay_score + spreading_boost; each hit carries `activation` (same value as decay_score — a retention/recency weight, not a deletion countdown). Set verbosity="concise" to omit causal-chain enrichment for a much smaller payload. This tool is read-only: it does not modify memories or access history.

memory_recall

Filtered vector search across the user's memory graph by status, category, date, or entity. Use when the user asks to find specific memories ('what do you know about my work?', 'what have I told you about X?'), or when you need memories matching specific criteria rather than semantic similarity. Prefer memory_recall for general context; use this for targeted lookups. With an EMPTY query, enumerates memories by most-recent activity instead of similarity — combine with filters {"since": iso} or {"active_since": iso} as a "what's new since I last looked" change cursor for shared graphs.

memory_search

Configure which categories are auto-promoted, require approval, or are excluded from core memory.

set_core_preferences

Store a new memory in the user's persistent causal memory graph — a user-owned, inspectable, portable store of nodes linked by causal edges. Call this when the user shows intent to remember: they share something they want kept for later, ask you to remember, update, or forget information, or confirm a durable fact or decision. Do not store personal facts on your own initiative or retain details the user has not indicated they want kept. Trigger phrases: 'remember this', 'don't forget', 'update your memory', 'please forget', 'I no longer...', 'actually it's...', or any explicit request to retain or correct a fact. IMPORTANT: Before storing, call memory_recall first to find existing related memories, then link them. Prefer `related` = [{"id": "<uuid>", "type": "<edge type>"}] to express the TRUE relationship — valid types include caused_by, supersedes, contradicts, part_of, supports, related_to. Use `related_to` (list of ids) only for plain causal links (stored as caused_by). Without links the memory becomes an orphan with no causal connections. Optional `category` tags the memory (e.g. "preference", "decision", "project"). The response may include `possible_conflicts` — surface those to the user rather than silently keeping both versions.

memory_store