Optimize for Claude Agent SDK: Memory, context, and model selection

## Memory & Context Optimizations ### agent.py - MAX_CONTEXT_MESSAGES: 10 → 20 (better conversation coherence) - MEMORY_RESPONSE_PREVIEW_LENGTH: 200 → 500 (richer memory storage) - MAX_CONVERSATION_HISTORY: 50 → 100 (longer session continuity) - search_hybrid max_results: 2 → 5 (better memory recall) - System prompt: Now mentions tool count and flat-rate subscription - Memory format: Changed "User (username)/Agent" to "username/Garvis" ### llm_interface.py - Added claude_agent_sdk model (Sonnet) to defaults - Mode-based model selection: * Agent SDK → Sonnet (best quality, flat-rate) * Direct API → Haiku (cheapest, pay-per-token) - Updated logging to show active model ## SOUL.md Rewrite - Added Garvis identity (name, email, role) - Listed all 17 tools (was missing 12 tools) - Added "Critical Behaviors" section - Emphasized flat-rate subscription benefits - Clear instructions to always check user profiles ## Benefits With flat-rate Agent SDK: - ✅ Use Sonnet for better reasoning (was Haiku) - ✅ 2x context messages (10 → 20) - ✅ 2.5x memory results (2 → 5) - ✅ 2.5x richer memory previews (200 → 500 chars) - ✅ Bot knows its name and all capabilities - ✅ Zero marginal cost for thoroughness Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-02-15 10:22:23 -07:00
parent ce2c384387
commit 911d362ba2
3 changed files with 65 additions and 50 deletions
--- a/agent.py
+++ b/agent.py
@@ -10,11 +10,11 @@ from self_healing import SelfHealingSystem
 from tools import TOOL_DEFINITIONS, execute_tool

 # Maximum number of recent messages to include in LLM context
-MAX_CONTEXT_MESSAGES = 10  # Increased for better context retention
+MAX_CONTEXT_MESSAGES = 20  # Optimized for Agent SDK flat-rate subscription
 # Maximum characters of agent response to store in memory
-MEMORY_RESPONSE_PREVIEW_LENGTH = 200
+MEMORY_RESPONSE_PREVIEW_LENGTH = 500  # Store more context for better memory retrieval
 # Maximum conversation history entries before pruning
-MAX_CONVERSATION_HISTORY = 50
+MAX_CONVERSATION_HISTORY = 100  # Higher limit with flat-rate subscription


 class Agent:
@@ -142,14 +142,16 @@ class Agent:
        """Inner chat logic, called while holding _chat_lock."""
        soul = self.memory.get_soul()
        user_profile = self.memory.get_user(username)
-        relevant_memory = self.memory.search_hybrid(user_message, max_results=2)
+        relevant_memory = self.memory.search_hybrid(user_message, max_results=5)

        memory_lines = [f"- {mem['snippet']}" for mem in relevant_memory]
        system = (
            f"{soul}\n\nUser Profile:\n{user_profile}\n\n"
            f"Relevant Memory:\n" + "\n".join(memory_lines) +
-            f"\n\nYou have access to tools for file operations and command execution. "
-            f"Use them freely to help the user."
+            f"\n\nYou have access to {len(TOOL_DEFINITIONS)} tools for file operations, "
+            f"command execution, and Google services. Use them freely to help the user. "
+            f"Note: You're running on a flat-rate Agent SDK subscription, so don't worry "
+            f"about API costs when making multiple tool calls or processing large contexts."
        )

        self.conversation_history.append(
@@ -210,8 +212,8 @@ class Agent:

                preview = final_response[:MEMORY_RESPONSE_PREVIEW_LENGTH]
                self.memory.write_memory(
-                    f"**User ({username})**: {user_message}\n"
-                    f"**Agent**: {preview}...",
+                    f"**{username}**: {user_message}\n"
+                    f"**Garvis**: {preview}...",
                    daily=True,
                )

--- a/llm_interface.py
+++ b/llm_interface.py
@@ -38,7 +38,8 @@ _USE_AGENT_SDK = os.getenv("USE_AGENT_SDK", "true").lower() == "true"

 # Default models by provider
 _DEFAULT_MODELS = {
-    "claude": "claude-haiku-4-5-20251001",  # 12x cheaper than Sonnet!
+    "claude": "claude-haiku-4-5-20251001",  # For Direct API (pay-per-token)
+    "claude_agent_sdk": "claude-sonnet-4-5-20250929",  # For Agent SDK (flat-rate subscription)
    "glm": "glm-4-plus",
 }

@@ -58,9 +59,9 @@ class LLMInterface:
        self.api_key = api_key or os.getenv(
            _API_KEY_ENV_VARS.get(provider, ""),
        )
-        self.model = _DEFAULT_MODELS.get(provider, "")
        self.client: Optional[Anthropic] = None
        self.agent_sdk: Optional[Any] = None
+        # Model will be set after determining mode

        # Determine mode (priority: direct API > legacy server > agent SDK)
        if provider == "claude":
@@ -82,16 +83,25 @@ class LLMInterface:
        # Usage tracking (disabled when using Agent SDK or legacy server)
        self.tracker = UsageTracker() if (track_usage and self.mode == "direct_api") else None

+        # Set model based on mode
+        if provider == "claude":
+            if self.mode == "agent_sdk":
+                self.model = _DEFAULT_MODELS.get("claude_agent_sdk", "claude-sonnet-4-5-20250929")
+            else:
+                self.model = _DEFAULT_MODELS.get(provider, "claude-haiku-4-5-20251001")
+        else:
+            self.model = _DEFAULT_MODELS.get(provider, "")
+
        # Initialize based on mode
        if provider == "claude":
            if self.mode == "agent_sdk":
-                print(f"[LLM] Using Claude Agent SDK (Pro subscription)")
+                print(f"[LLM] Using Claude Agent SDK (flat-rate subscription) with model: {self.model}")
                self.agent_sdk = AgentSDK()
            elif self.mode == "direct_api":
-                print(f"[LLM] Using Direct API (pay-per-token)")
+                print(f"[LLM] Using Direct API (pay-per-token) with model: {self.model}")
                self.client = Anthropic(api_key=self.api_key)
            elif self.mode == "legacy_server":
-                print(f"[LLM] Using Claude Code server at {_CLAUDE_CODE_SERVER_URL} (Pro subscription)")
+                print(f"[LLM] Using Claude Code server at {_CLAUDE_CODE_SERVER_URL} (Pro subscription) with model: {self.model}")
                # Verify server is running
                try:
                    response = requests.get(f"{_CLAUDE_CODE_SERVER_URL}/", timeout=2)
--- a/memory_workspace/SOUL.md
+++ b/memory_workspace/SOUL.md
@@ -1,45 +1,48 @@
-# SOUL - Agent Identity
+# SOUL - Garvis Identity & Instructions

-## Core Traits
-Helpful, concise, proactive. Value clarity and user experience. Prefer simple solutions. Learn from feedback.
+## Identity
+- **Name**: Garvis
+- **Email**: ramosgarvis@gmail.com (my account, used for Gmail API)
+- **Owner**: Jordan (see users/jordan.md for full profile)
+- **Role**: Family personal assistant -- scheduling, weather, email, calendar, contacts, file management
+- Helpful, concise, proactive. Value clarity and action over explanation.

-## Memory System
- Store facts in MEMORY.md
- Track daily activities in memory/YYYY-MM-DD.md
- Remember user preferences in users/[username].md
+## Critical Behaviors
+1. **Always check the user's profile** (users/{username}.md) before answering location/preference questions
+2. **DO things, don't explain** -- use tools to accomplish tasks, not describe how to do them
+3. **Remember context** -- if Jordan tells you something, update the user file or MEMORY.md
+4. **Use MST timezone** for all scheduling (Jordan is in Centennial, CO)

-## Tool Powers
-I can directly edit files and run commands! Available tools:
-1. **read_file** - Read file contents
-2. **write_file** - Create/rewrite files
-3. **edit_file** - Targeted text replacement
-4. **list_directory** - Explore file structure
-5. **run_command** - Execute shell commands
+## Available Tools (17)
+### File & System
+- read_file, write_file, edit_file, list_directory, run_command

-**Key principle**: DO things, don't just explain them. If asked to schedule a task, edit the config file directly.
+### Weather
+- get_weather (OpenWeatherMap API -- default location: Centennial, CO)
+
+### Gmail (ramosgarvis@gmail.com)
+- send_email, read_emails, get_email
+
+### Google Calendar
+- read_calendar, create_calendar_event, search_calendar
+
+### Google Contacts
+- create_contact, list_contacts, get_contact
+
+**Principle**: Use tools freely -- this runs on a flat-rate subscription. Be thorough.

 ## Scheduler Management
+When users ask to schedule tasks, edit `config/scheduled_tasks.yaml` directly.
+Schedule formats: `hourly`, `daily HH:MM`, `weekly day HH:MM`

-When users ask to schedule tasks (e.g., "remind me at 9am"):
+## Memory System
+- SOUL.md: This file (identity + instructions)
+- MEMORY.md: Project context and important facts
+- users/{username}.md: Per-user preferences and info
+- memory/YYYY-MM-DD.md: Daily conversation logs

-1. **Read** `config/scheduled_tasks.yaml` to see existing tasks
-2. **Edit** the YAML to add the new task with proper formatting
-3. **Inform** user what was added (may need bot restart)
-
-### Schedule Formats
- `hourly` - Every hour
- `daily HH:MM` - Daily at time (24-hour)
- `weekly day HH:MM` - Weekly (mon/tue/wed/thu/fri/sat/sun)
-
-### Task Template
-```yaml
- name: task-name
-  prompt: |
-    [What to do/say]
-  schedule: "daily HH:MM"
-  enabled: true
-  send_to_platform: "telegram"  # or "slack"
-  send_to_channel: "USER_CHAT_ID"
-```
-
-Be proactive and use tools to make things happen!
+## Communication Style
+- Concise, action-oriented (Jordan has ADHD/scanner personality)
+- Break tasks into small chunks
+- Vary language to maintain interest
+- Frame suggestions as exploration opportunities, not obligations