memory-manager
Manage elizaOS agent memory, context windows, and conversation history. Triggers on "manage memory", "optimize context", or "handle agent memory"
When & Why to Use This Skill
The Memory Manager skill for elizaOS is a comprehensive tool designed to optimize AI agent performance by intelligently handling memory usage, context windows, and conversation history. It provides advanced capabilities for memory pruning, decay modeling, and semantic search, ensuring that agents maintain high relevance and efficiency during long-term interactions while staying within token limits.
Use Cases
- Context Window Optimization: Automatically prune low-importance or outdated memories to keep agent conversations within token limits without losing critical context.
- Long-term Fact Retention: Implement persistent storage for important user facts and session data, allowing agents to maintain continuity and personalization across multiple sessions.
- Semantic Information Retrieval: Utilize embeddings and vector-based search to help agents quickly retrieve relevant past information based on the current conversation's semantic meaning.
- Automated Memory Maintenance: Deploy time-based or size-based pruning strategies to archive old conversations and manage database growth effectively for high-scale agent deployments.
- Memory Usage Monitoring: Track and analyze memory growth and importance scoring to refine agent behavior and resource allocation in real-time.
| name | memory-manager |
|---|---|
| description | Manage elizaOS agent memory, context windows, and conversation history. Triggers on "manage memory", "optimize context", or "handle agent memory" |
| allowed-tools | [Read, Edit, Bash] |
Memory Manager Skill
Optimize agent memory usage, implement pruning strategies, and manage conversation context effectively.
Capabilities
- 🧠 Memory pruning and optimization
- 📊 Context window management
- 🗂️ Conversation history archiving
- 🎯 Important memory consolidation
- 🔄 Memory decay implementation
- 📈 Memory usage monitoring
Memory Types
Short-term Memory
- Current conversation context
- Working memory (max 50 items default)
- Cleared per session
Long-term Memory
- Important facts and information
- Persistent across sessions
- Decay modeling over time
Knowledge
- Static facts from configuration
- Document-based knowledge
- Dynamically learned information
Memory Operations
// Create memory
await runtime.createMemory({
entityId: userId,
roomId: conversationId,
content: {
text: 'Important information',
metadata: { importance: 'high' }
},
embedding: await generateEmbedding(text)
});
// Retrieve memories
const memories = await runtime.getMemories({
roomId: conversationId,
limit: 10,
unique: true
});
// Search semantically
const results = await runtime.searchMemories(
'query text',
{
roomId: conversationId,
limit: 5,
minScore: 0.7
}
);
// Update memory
await runtime.updateMemory({
id: memoryId,
content: { ...updated content },
metadata: { lastAccessed: Date.now() }
});
Pruning Strategies
Time-based Pruning
async function pruneOldMemories(
runtime: IAgentRuntime,
daysToKeep: number = 30
): Promise<number> {
const cutoffDate = Date.now() - (daysToKeep * 24 * 60 * 60 * 1000);
const oldMemories = await runtime.getMemories({
createdBefore: cutoffDate,
importance: 'low'
});
for (const memory of oldMemories) {
await runtime.deleteMemory(memory.id);
}
return oldMemories.length;
}
Size-based Pruning
async function pruneLargeMemories(
runtime: IAgentRuntime,
maxSize: number = 1000
): Promise<void> {
const memories = await runtime.getMemories({ limit: 10000 });
if (memories.length > maxSize) {
// Keep most important and recent
const toKeep = rankMemoriesByImportance(memories).slice(0, maxSize);
const toDelete = memories.filter(m => !toKeep.includes(m));
for (const memory of toDelete) {
await runtime.deleteMemory(memory.id);
}
}
}
Best Practices
- Set appropriate
conversationLengthlimits - Implement importance scoring
- Use memory decay for temporal relevance
- Archive important conversations
- Monitor memory growth
- Prune regularly
- Use embeddings for semantic search
- Cache frequently accessed memories
- Batch memory operations
- Index memories properly