跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
设计与 UI

memory-management

Design and operate memory systems for long-running AI agents. Covers context window optimization, summarization strategies, vector-based retrieval, episodic memory, memory consolidation, and garbage collection for production agent systems.

cosmicstack-labs · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add cosmicstack-labs/mercury-agent-skills --skill memory-management

技能说明


name: memory-management description: 'Design and operate memory systems for long-running AI agents. Covers context window optimization, summarization strategies, vector-based retrieval, episodic memory, memory consolidation, and garbage collection for production agent systems.' metadata: author: cosmicstack-labs version: 1.0.0 category: ai-ml tags: - memory-management - context-window - vector-database - summarization - rag - long-running-agents - memory-consolidation

Memory Management for Long-Running Agents

Overview

Long-running agents face a fundamental problem: they can't remember everything, but forgetting the wrong thing breaks their usefulness. This skill covers memory architectures that balance context retention, token budget, and retrieval accuracy for agents that run for hours, days, or continuously.


Core Concepts

The Memory Problem

IssueSymptomCost
Context OverflowAgent forgets early instructionsTask failure, incoherent responses
Token BloatEvery message keeps growing10x+ cost increase per task
Memory PollutionIrrelevant memories distract agentHallucination, off-target responses
Stale MemoriesOutdated information used as factIncorrect decisions
Memory LeaksUnused data accumulates unboundedCrash from OOM, endless context

Memory Tiers

TierStorageCapacityAccess SpeedCostBest For
L1 — WorkingIn-context (LLM window)8K-200K tokensInstant$$$Current task, immediate context
L2 — RecentSliding window buffer~2K turns< 10ms$$Recent conversation history
L3 — EpisodicEvent log / timeseriesMillions of events< 50ms$Past actions, outcomes, decisions
L4 — SemanticVector databaseUnlimited< 100ms$Knowledge, facts, relationships
L5 — ArchivalObject storageUnlimited> 1s$Backups, compliance, audit

Step-by-Step Implementation

Step 1: Build a Tiered Memory System

from dataclasses import dataclass, field
from typing import Optional
import json
import time

@dataclass
class MemoryEntry:
    content: str
    timestamp: float = None
    importance: float = 0.5  # 0.0 (trivial) to 1.0 (critical)
    tags: list[str] = field(default_factory=list)
    token_count: int = 0
    
    def __post_init__(self):
        if self.timestamp is None:
            self.timestamp = time.time()

class TieredMemory:
    """Multi-tier memory with automatic promotion and demotion."""
    
    def __init__(self, llm, vector_store, max_context_tokens: int = 8000):
        self.llm = llm
        self.vector_store = vector_store
        self.max_context_tokens = max_context_tokens
        
        # L1: Working context (in-memory)
        self.working_memory: list[MemoryEntry] = []
        self.current_tokens = 0
        
        # L2: Recent history buffer
        self.recent_buffer: list[MemoryEntry] = []
        self.buffer_size = 50
        
        # L3: Episodic memory
        self.episodes: list[MemoryEntry] = []
        
        # L4: Semantic memory (vector DB)
        # Initialized externally
    
    async def remember(self, content: str, importance: float = 0.5, 
                       tags: list[str] = None):
        """Store a new memory across tiers."""
        entry = MemoryEntry(
            content=content,
            importance=importance,
            tags=tags or [],
            token_count=self._count_tokens(content)
        )
        
        # Always add to working memory
        self.working_memory.append(entry)
        self.current_tokens += entry.token_count
        
        # If important, store in episodic + semantic
        if importance > 0.7:
            self.episodes.append(entry)
            await self.vector_store.store(entry)
        
        # Trim if needed
        await self._trim_working_memory()

Step 2: Implement Context Window Management

class ContextManager:
    """Optimize what stays in the context window."""
    
    def __init__(self, tiered_memory: TieredMemory, 
                 summarizer, max_tokens: int = 8000):
        self.memory = tiered_memory
        self.summarizer = summarizer
        self.max_tokens = max_tokens
        self.reserved_tokens = 2000  # Reserve for new input/output
    
    async def build_context(self, task: str, top_k: int = 5) -> list[dict]:
        """Build the optimal context for a task."""
        
        available_tokens = self.max_tokens - self.reserved_tokens
        
        # 1. Start with high-importance working memory
        context = []
        tokens_used = 0
        
        working = sorted(
            self.memory.working_memory,
            key=lambda e: e.importance,
            reverse=True
        )
        
        for entry in working:
            if tokens_used + entry.token_count > available_tokens:
                break
            context.append({"role": "system", "content": entry.content})
            tokens_used += entry.token_count
        
        # 2. Add semantically relevant memories
        relevant = await self.memory.vector_store.search(task, k=top_k)
        for mem in relevant:
            if tokens_used + mem.token_count > available_tokens:
                break
            context.append({"role": "system", "content": mem.content})
            tokens_used += mem.token_count
        
        # 3. If we had to drop items, add a summary
        if len(context) < len(working):
            summary = await self._get_summary()
            context.insert(0, {"role": "system", 
                               "content": f"[Summary of earlier context]: {summary}"})
        
        return context
    
    async def _get_summary(self) -> str:
        """Summarize what was excluded from context."""
        excluded =
阅读完整来源 ↗