advanced⏱70 minLección 4 de 10

Context Management

Master context windows, memory systems, retrieval-augmented generation, summarization strategies, and efficient context handling for AI agents.

Context Management

The Context Challenge

AI agents face a fundamental constraint: limited context windows. Even the largest models (200K+ tokens) cannot hold an entire codebase, conversation history, and task instructions simultaneously. Effective context management is the difference between a capable agent and one that repeatedly loses track of important information.

💡 Tip: Use Ctrl + Mouse Wheel to zoom, or click the fullscreen icon

ℹ️Note

Context management is not just about fitting things into a token budget. It's about making intelligent decisions about what to keep, what to compress, what to fetch on demand, and what to forget.

Context Window Architecture

Understanding how different context types interact within the window is crucial for efficient agent design.

💡 Tip: Use Ctrl + Mouse Wheel to zoom, or click the fullscreen icon

Memory Systems

Effective agents use a tiered memory architecture to manage context efficiently:

python

class TieredMemory:
    def __init__(self):
        # Ephemeral: current conversation turn
        self.working_memory = {}
        # Short-term: recent conversation (sliding window)
        self.short_term = []
        # Long-term: persistent facts across sessions
        self.long_term = {}
        # Semantic: indexed knowledge for retrieval
        self.semantic_store = []

    def add_to_working(self, key, value):
        self.working_memory[key] = value

    def get_working(self, key):
        return self.working_memory.get(key)

    def add_to_short_term(self, message, max_size=20):
        self.short_term.append(message)
        if len(self.short_term) > max_size:
            # Evict oldest messages
            self.short_term = self.short_term[-max_size:]

    def remember(self, key, value, importance=1.0):
        # High importance facts persist longer
        self.long_term[key] = {
            "value": value,
            "importance": importance,
            "timestamp": __import__('time').time()
        }

    def recall(self, key):
        entry = self.long_term.get(key)
        if entry:
            return entry["value"]
        return None

    def recall_by_relevance(self, query, top_k=3):
        # Simple keyword matching for semantic recall
        scores = []
        for key, entry in self.long_term.items():
            if query.lower() in key.lower():
                scores.append((entry["importance"], key, entry["value"]))
        scores.sort(reverse=True)
        return [(k, v) for _, k, v in scores[:top_k]]

    def store_semantic(self, text, metadata=None):
        self.semantic_store.append({
            "text": text,
            "metadata": metadata or {},
            "timestamp": __import__('time').time()
        })

    def summarize_and_archive(self):
        """Compress short-term memory into a summary for long-term storage."""
        if len(self.short_term) > 10:
            texts = [m["content"] for m in self.short_term[:5]]
            summary = " | ".join(texts)
            self.long_term["conversation_summary"] = {
                "value": summary[:500],
                "importance": 0.5,
                "timestamp": __import__('time').time()
            }
            # Keep only recent messages
            self.short_term = self.short_term[-5:]


memory = TieredMemory()

# Store project facts long-term
memory.remember("project_name", "Nova Platform", importance=1.0)
memory.remember("tech_stack", ["Python", "FastAPI", "React", "PostgreSQL"], importance=0.9)
memory.remember("deployment_env", "production", importance=0.8)

# Add conversation to short-term
memory.add_to_short_term({"role": "user", "content": "Fix the login bug"})
memory.add_to_short_term({"role": "assistant", "content": "Found the issue in auth.py"})

# Retrieve relevant facts
relevant = memory.recall_by_relevance("deploy")
print(f"Deployment context: {relevant}")

# Summarize when needed
memory.summarize_and_archive()

Retrieval-Augmented Generation (RAG)

RAG enables agents to access large knowledge bases by fetching only the most relevant information at query time.

python

class RAGSystem:
    def __init__(self):
        self.documents = []
        self.index = {}

    def add_document(self, doc_id, content, metadata=None):
        self.documents.append({
            "id": doc_id,
            "content": content,
            "metadata": metadata or {}
        })
        # Simple keyword indexing
        for word in content.lower().split():
            if word not in self.index:
                self.index[word] = []
            self.index[word].append(len(self.documents) - 1)

    def search(self, query, top_k=3):
        query_words = query.lower().split()
        scores = {}

        for q_word in query_words:
            for doc_idx in self.index.get(q_word, []):
                if doc_idx not in scores:
                    scores[doc_idx] = 0
                scores[doc_idx] += 1

        # Sort by relevance score
        ranked = sorted(scores.items(), key=lambda x: x[1], reverse=True)
        results = []
        for doc_idx, score in ranked[:top_k]:
            doc = self.documents[doc_idx]
            results.append({
                "id": doc["id"],
                "content": doc["content"][:500],  # Truncate for context
                "relevance": score / len(query_words),
                "metadata": doc["metadata"]
            })
        return results

    def build_context_prompt(self, query, max_tokens=2000):
        results = self.search(query)
        context_parts = []
        token_count = 0

        for r in results:
            snippet = f"[Source: {r['id']}]\n{r['content']}\n"
            snippet_tokens = len(snippet.split())
            if token_count + snippet_tokens > max_tokens:
                break
            context_parts.append(snippet)
            token_count += snippet_tokens

        return "\n---\n".join(context_parts)


# Build a RAG system for an agent
rag = RAGSystem()
rag.add_document(
    "auth-flow.md",
    "The authentication flow uses JWT tokens with 1-hour expiry. "
    "Refresh tokens are stored in HTTP-only cookies and have 7-day expiry. "
    "OAuth2 is supported for Google and GitHub providers.",
    {"type": "documentation", "module": "auth"}
)
rag.add_document(
    "api-endpoints.md",
    "POST /api/auth/login - Authenticate user\n"
    "POST /api/auth/refresh - Refresh access token\n"
    "POST /api/auth/logout - Invalidate session\n"
    "GET /api/auth/profile - Get user profile",
    {"type": "api", "module": "auth"}
)
rag.add_document(
    "db-schema.sql",
    "CREATE TABLE users (id UUID PRIMARY KEY, email VARCHAR UNIQUE, "
    "password_hash VARCHAR, created_at TIMESTAMP);\n"
    "CREATE TABLE sessions (id UUID PRIMARY KEY, user_id UUID REFERENCES users(id), "
    "token VARCHAR, expires_at TIMESTAMP);",
    {"type": "schema", "module": "database"}
)

context = rag.build_context_prompt("How does authentication work?")
print(context)

⚠️Warning

RAG is not magic. The quality of retrieval depends on the quality of your indexing strategy. Simple keyword search works for small codebases, but production systems need embeddings-based semantic search. Always validate that the retriever finds the right documents for common queries.

Summarization Strategies

When context windows fill up, summarization compresses information while preserving key details.

python

class SummarizationEngine:
    def __init__(self, llm_client):
        self.llm = llm_client

    def summarize_conversation(self, messages):
        """Summarize a conversation history into key points."""
        if len(messages) < 3:
            return messages  # Too short to summarize

        text = "\n".join(
            f"{m['role']}: {m['content']}" for m in messages
        )

        prompt = f"""
        Summarize the following conversation, preserving:
        1. All decisions made
        2. Key facts discovered
        3. Action items or next steps
        4. Technical details (file paths, function names, versions)

        Conversation:
        {text}

        Summary:
        """
        return self.llm.complete(prompt)

    def hierarchical_summary(self, sessions, max_levels=3):
        """Create a hierarchical summary of multiple sessions."""
        summaries = []

        for session in sessions:
            summary = self.summarize_conversation(session)
            summaries.append(summary)

        # If we have too many session summaries, summarize those too
        level = 1
        while len(summaries) > 5 and level < max_levels:
            batch_size = 3
            new_summaries = []
            for i in range(0, len(summaries), batch_size):
                batch = summaries[i:i + batch_size]
                combined = "\n".join(batch)
                meta_summary = self.llm.complete(
                    f"Summarize these session summaries:\n{combined}"
                )
                new_summaries.append(meta_summary)
            summaries = new_summaries
            level += 1

        return summaries

    def token_aware_truncation(self, text, max_tokens=4000):
        """Truncate text intelligently at token boundaries."""
        tokens = text.split()
        if len(tokens) <= max_tokens:
            return text

        # Keep beginning and end, summarize middle
        keep_start = max_tokens // 3
        keep_end = max_tokens // 3
        middle_start = text[:keep_start * 10]  # Approximate char→token
        middle_end = text[-keep_end * 10:]

        return f"{middle_start}\n\n...[{len(tokens) - keep_start - keep_end} tokens omitted]...\n\n{middle_end}"

yaml

# Summarization configuration for an agent
summarization:
  strategy: hierarchical
  thresholds:
    short_term: 10  # Summarize when >10 messages
    session: 50      # Summarize when >50 messages per session
    total: 200       # Create meta-summary across sessions
  include_in_summary:
    - decisions_made
    - code_changes
    - error_patterns
    - user_preferences
  preserve_fields:
    - file_paths
    - function_names
    - version_numbers
    - api_endpoints

💡Tip

Use hierarchical summarization for long-running agent sessions. Summarize at three levels: (1) individual conversation turns, (2) session-level summaries, (3) cross-session meta-summaries. This preserves critical information while maintaining a compact context.

Context Prioritization

Not all context is equally important. Agents should prioritize what goes into the limited context window.

python

class ContextPrioritizer:
    def __init__(self, max_tokens=128000):
        self.max_tokens = max_tokens
        self.priority_levels = {
            "critical": 5,
            "high": 4,
            "medium": 3,
            "low": 2,
            "background": 1
        }

    def prioritize(self, context_items):
        """Sort and filter context items by priority, fit within token budget."""
        # Assign priority scores
        scored = []
        for item in context_items:
            score = self._calculate_priority(item)
            tokens = len(item["content"].split())
            scored.append((score, tokens, item))

        # Sort by priority (descending)
        scored.sort(key=lambda x: x[0], reverse=True)

        # Fit within budget
        result = []
        used_tokens = 0

        # Always include system prompt and current request
        for item in context_items:
            if item["type"] in ("system_prompt", "current_request"):
                result.append(item)
                used_tokens += len(item["content"].split())

        # Add high priority items
        remaining = self.max_tokens - used_tokens
        for score, tokens, item in scored:
            if item["type"] in ("system_prompt", "current_request"):
                continue  # Already included
            if used_tokens + tokens <= self.max_tokens:
                result.append(item)
                used_tokens += tokens

        return result

    def _calculate_priority(self, item):
        base = self.priority_levels.get(item.get("priority", "medium"), 3)

        # Boost relevance to current task
        if item.get("relevance_to_current", 0) > 0.8:
            base += 1
        # Boost recent items
        if item.get("recency", 0) > 0.9:
            base += 1
        # Penalize very large items
        tokens = len(item["content"].split())
        if tokens > 10000:
            base -= 1

        return max(1, base)


prioritizer = ContextPrioritizer(max_tokens=8000)
context = [
    {"type": "system_prompt", "content": "You are a coding agent...", "priority": "critical"},
    {"type": "current_request", "content": "Fix the payment bug", "priority": "critical"},
    {"type": "file_content", "content": "def process(): ...", "priority": "high", "relevance_to_current": 0.9},
    {"type": "conversation_history", "content": "User: The payment fails...", "priority": "medium", "recency": 1.0},
    {"type": "project_readme", "content": "# Project...", "priority": "low"},
]
prioritized = prioritizer.prioritize(context)
print(f"Included {len(prioritized)} items in context window")

Context Window Limits by Model

Model Provider	Model	Context Window	Effective Budget
OpenAI	GPT-4o	128K tokens	~96K (after overhead)
OpenAI	GPT-4o-mini	128K tokens	~96K
Anthropic	Claude Sonnet 4	200K tokens	~170K
Anthropic	Claude Opus 4	200K tokens	~170K
Google	Gemini 1.5 Pro	1M tokens	~800K
Meta	Llama 3.1 70B	128K tokens	~100K
Mistral	Mixtral 8x22B	64K tokens	~48K

ℹ️Note

The "Effective Budget" accounts for the tokens consumed by the model's internal processing (system prompts, tokenization overhead, response generation). A 128K window typically has about 96K tokens available for your content.

Practical Context Management in OpenCode

OpenCode provides built-in context management through its agent configuration system:

json

{
  "agents": {
    "default": {
      "model": "gpt-4o",
      "description": "General coding agent",
      "context": {
        "maxTokens": 128000,
        "summarizationThreshold": 80000,
        "memoryStrategy": "tiered",
        "shortTermLimit": 20,
        "longTermEnabled": true,
        "ragEnabled": true,
        "ragMaxDocs": 5,
        "prioritizeByRelevance": true
      }
    },
    "code-reviewer": {
      "model": "claude-sonnet-4-20250514",
      "description": "Code review specialist",
      "context": {
        "maxTokens": 200000,
        "summarizationThreshold": 150000,
        "focusOnCurrentChanges": true,
        "includeGitDiff": true
      }
    }
  }
}

yaml

# context-management-strategy.yaml
strategy:
  name: "tiered-context"
  layers:
    working:
      label: "Working Memory"
      size: "2-3 turns"
      content: "Current task, immediate tool results"
      eviction: "FIFO"
    short_term:
      label: "Short-Term Memory"
      size: "15-20 turns"
      content: "Recent conversation, recent file reads"
      eviction: "Summarize and archive"
    long_term:
      label: "Long-Term Memory"
      size: "Unlimited (external)"
      content: "Project facts, decisions, user preferences"
      access: "On-demand retrieval"
    episodic:
      label: "Episodic Memory"
      size: "Compressed"
      content: "Key events, errors encountered, solutions found"
      access: "Similarity search"

Common Context Management Pitfalls

Pitfall	Symptom	Solution
Context overflow	Agent loses track of early instructions	Implement sliding window + summarization
Stale context	Agent references outdated information	Add recency scoring in prioritization
Missing context	Agent asks for information it already has	Improve long-term memory recall
Token waste	Agent includes irrelevant file contents	Use RAG with precise queries
Memory explosion	Context grows unboundedly	Set hard limits with summarization triggers
Fact drift	Agent contradicts itself across sessions	Use persistent long-term memory store

⚠️Warning

The most expensive mistake is losing critical information because of poor context management. An agent that forgets your project's tech stack or a key decision will produce inconsistent results. Always validate that your memory system preserves essential facts.

Practice Exercises

Practice Question

What is the primary purpose of a sliding window in context management?

Practice Question

In a tiered memory system, what type of information should be stored in long-term memory?

Practice Question

What happens when an agent's context window exceeds the model's token limit?

Practice Question

What is RAG (Retrieval-Augmented Generation) used for in agent context management?

Practice Question

An agent has a 128K token limit and is using about 100K tokens for conversation history. What strategy should it use when the user sends a new, complex request?

Practice Question

What is the 'effective budget' of a 128K token context window?

Practice Question

What is the main risk of using only a sliding window without summarization?

Practice Question

In context prioritization, which items should always be included in the context window regardless of token budget?

✅**Key Takeaways**

Context management is about intelligent decisions on what to keep, compress, fetch, and forget
Tiered memory (working, short-term, long-term, semantic) provides a scalable architecture
RAG enables access to large knowledge bases without filling the context window
Summarization strategies (linear, hierarchical, token-aware) compress information effectively
Context prioritization ensures critical items always fit within the token budget
Different models have different context limits; design for the effective budget
OpenCode provides built-in context configuration for agents
Without explicit management, context overflow silently degrades agent performance
Common pitfalls include context overflow, stale context, missing context, and memory explosion

Progreso40%

Skills & Capabilities

Agent Pipelines

Contenido del Curso

1. Introduction to AI Agents 2. Agent Architecture 3. Skills & Capabilities 4. Context Management 5. Agent Pipelines 6. OpenCode as Development Tool 7. Building Custom Agent Skills 8. Prompt Engineering for Agents 9. Best Practices & Patterns 10. Real-World Agent Projects