Added llama.cpp and implemented long term memory wrt to TITANS Paper

Author: vishal veerareddy

.env

@@ -1,7 +1,7 @@
 
 # Primary model provider
-# Options: ollama, openrouter, azure-openai,openai, azure-anthropic, databricks
-MODEL_PROVIDER=azure-openai
+# Options: ollama, openrouter, azure-openai, openai, azure-anthropic, databricks, llamacpp
+MODEL_PROVIDER=llamacpp
 
 # Ollama endpoint (local by default)
 OLLAMA_ENDPOINT=http://localhost:11434
@@ -80,11 +80,29 @@ AZURE_OPENAI_DEPLOYMENT=Kimi-K2-Thinking
 
 # OpenAI organization ID (optional, for organization-level API keys)
 # OPENAI_ORGANIZATION=org-your-org-id
+
+# ==============================================================================
+# llama.cpp Configuration (Local GGUF Models)
+# ==============================================================================
+
+# llama.cpp server endpoint (default: http://localhost:8080)
+# Start with: ./llama-server -m model.gguf --port 8080
+LLAMACPP_ENDPOINT=http://localhost:8080
+
+# Model name (for logging purposes, llama.cpp uses the loaded model)
+LLAMACPP_MODEL=default
+
+# Request timeout in milliseconds (default: 120000)
+LLAMACPP_TIMEOUT_MS=120000
+
+# Optional API key (for secured llama.cpp servers)
+# LLAMACPP_API_KEY=
+
 # ==============================================================================
 # Server Configuration
 # ==============================================================================
 
-PORT=8080
+PORT=8081
 LOG_LEVEL=debug  # Use 'debug' for detailed logs, 'info' for production
 WEB_SEARCH_ENDPOINT=http://localhost:8888/search
 
@@ -93,6 +111,62 @@ WEB_SEARCH_ENDPOINT=http://localhost:8888/search
 # - passthrough/client: Return tool calls to CLI for local execution
 TOOL_EXECUTION_MODE=client
 
+# ==============================================================================
+# Long-Term Memory System (Titans-Inspired)
+# ==============================================================================
+
+# Enable/disable the entire memory system
+# When enabled, automatically extracts and retrieves conversation memories
+# Default: true
+MEMORY_ENABLED=true
+
+# Maximum number of memories to inject into each request
+# Higher = more context but larger prompts
+# Default: 5, Range: 1-20
+MEMORY_RETRIEVAL_LIMIT=5
+
+# Minimum surprise score (0.0-1.0) required to store a memory
+# Filters out redundant information
+# Lower (0.1-0.2) = store more memories
+# Higher (0.4-0.5) = only store highly novel information
+# Default: 0.3
+MEMORY_SURPRISE_THRESHOLD=0.3
+
+# Auto-delete memories older than this many days
+# Prevents database bloat with stale information
+# Default: 90
+MEMORY_MAX_AGE_DAYS=90
+
+# Maximum total memories to keep (prunes least important when exceeded)
+# Default: 10000
+MEMORY_MAX_COUNT=10000
+
+# Enable importance decay over time (exponential decay)
+# Memories become less important as they age
+# Default: true
+MEMORY_DECAY_ENABLED=true
+
+# Days for importance to decay by 50% (exponential half-life)
+# Shorter (7-14) = prefer recent info
+# Longer (60-90) = value historical context
+# Default: 30
+MEMORY_DECAY_HALF_LIFE=30
+
+# Include memories with session_id=NULL (global memories) in all sessions
+# Global memories = project facts shared across all conversations
+# Default: true
+MEMORY_INCLUDE_GLOBAL=true
+
+# Where to inject memories in the prompt
+# Options: system (recommended), assistant_preamble
+# Default: system
+MEMORY_INJECTION_FORMAT=system
+
+# Enable automatic extraction of memories from assistant responses
+# Disable for read-only memory mode
+# Default: true
+MEMORY_EXTRACTION_ENABLED=true
+
 # ==============================================================================
 # Performance Tuning (Optional)
 # ==============================================================================

.env.example

@@ -6,7 +6,7 @@
 # ==============================================================================
 
 # Primary model provider to use
-# Options: databricks, azure-anthropic, azure-openai, openrouter, openai, ollama
+# Options: databricks, azure-anthropic, azure-openai, openrouter, openai, ollama, llamacpp
 # Default: databricks
 # MODEL_PROVIDER=databricks
 
@@ -22,7 +22,7 @@ PREFER_OLLAMA=true
 OLLAMA_MODEL=qwen2.5-coder:latest
 
 # Fallback provider when primary provider fails or for complex requests
-# Options: databricks, azure-anthropic, azure-openai, openrouter, openai
+# Options: databricks, azure-anthropic, azure-openai, openrouter, openai, llamacpp
 FALLBACK_PROVIDER=databricks
 
 # Enable automatic fallback (true = transparent fallback, false = fail on provider error)
@@ -84,6 +84,23 @@ OLLAMA_MAX_TOOLS_FOR_ROUTING=3
 # OpenAI organization ID (optional, for organization-level API keys)
 # OPENAI_ORGANIZATION=org-your-org-id
 
+# ==============================================================================
+# llama.cpp Configuration (Local GGUF Models)
+# ==============================================================================
+
+# llama.cpp server endpoint (default: http://localhost:8080)
+# Start with: ./llama-server -m model.gguf --port 8080
+# LLAMACPP_ENDPOINT=http://localhost:8080
+
+# Model name (for logging purposes, llama.cpp uses the loaded model)
+# LLAMACPP_MODEL=default
+
+# Request timeout in milliseconds (default: 120000)
+# LLAMACPP_TIMEOUT_MS=120000
+
+# Optional API key (for secured llama.cpp servers)
+# LLAMACPP_API_KEY=your-optional-api-key
+
 # ==============================================================================
 # Server Configuration
 # ==============================================================================
@@ -118,6 +135,62 @@ WEB_SEARCH_ENDPOINT=http://localhost:8888/search
 # Timeout per subagent (milliseconds)
 # AGENTS_TIMEOUT=120000
 
+# ==============================================================================
+# Long-Term Memory System (Titans-Inspired)
+# ==============================================================================
+
+# Enable/disable the entire memory system
+# When enabled, automatically extracts and retrieves conversation memories
+# Default: true
+# MEMORY_ENABLED=true
+
+# Maximum number of memories to inject into each request
+# Higher = more context but larger prompts
+# Default: 5, Range: 1-20
+# MEMORY_RETRIEVAL_LIMIT=5
+
+# Minimum surprise score (0.0-1.0) required to store a memory
+# Filters out redundant information
+# Lower (0.1-0.2) = store more memories
+# Higher (0.4-0.5) = only store highly novel information
+# Default: 0.3
+# MEMORY_SURPRISE_THRESHOLD=0.3
+
+# Auto-delete memories older than this many days
+# Prevents database bloat with stale information
+# Default: 90
+# MEMORY_MAX_AGE_DAYS=90
+
+# Maximum total memories to keep (prunes least important when exceeded)
+# Default: 10000
+# MEMORY_MAX_COUNT=10000
+
+# Enable importance decay over time (exponential decay)
+# Memories become less important as they age
+# Default: true
+# MEMORY_DECAY_ENABLED=true
+
+# Days for importance to decay by 50% (exponential half-life)
+# Shorter (7-14) = prefer recent info
+# Longer (60-90) = value historical context
+# Default: 30
+# MEMORY_DECAY_HALF_LIFE=30
+
+# Include memories with session_id=NULL (global memories) in all sessions
+# Global memories = project facts shared across all conversations
+# Default: true
+# MEMORY_INCLUDE_GLOBAL=true
+
+# Where to inject memories in the prompt
+# Options: system (recommended), assistant_preamble
+# Default: system
+# MEMORY_INJECTION_FORMAT=system
+
+# Enable automatic extraction of memories from assistant responses
+# Disable for read-only memory mode
+# Default: true
+# MEMORY_EXTRACTION_ENABLED=true
+
 # ==============================================================================
 # Performance & Security
 # ==============================================================================