Metadata-Version: 2.4
Name: agent-reminiscence
Version: 0.1.3
Summary: Hierarchical memory management for AI agents with vector search and graph relationships
Author: Agent Reminiscence Contributors
Maintainer: Agent Reminiscence Contributors
License: MIT
Project-URL: Homepage, https://github.com/Ganzzi/agent-reminiscence
Project-URL: Documentation, https://github.com/Ganzzi/agent-reminiscence/tree/main/docs
Project-URL: Repository, https://github.com/Ganzzi/agent-reminiscence
Project-URL: Issues, https://github.com/Ganzzi/agent-reminiscence/issues
Project-URL: Changelog, https://github.com/Ganzzi/agent-reminiscence/blob/main/CHANGELOG.md
Keywords: ai,memory,agents,vector-search,graph,pydantic-ai,llm,rag,knowledge-graph
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Developers
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Classifier: Topic :: Software Development :: Libraries :: Python Modules
Classifier: Framework :: AsyncIO
Requires-Python: >=3.10
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: pydantic>=2.0.0
Requires-Dist: pydantic-ai>=0.3.4
Requires-Dist: pydantic-ai-slim[anthropic,google,openai]>=0.3.4
Requires-Dist: psqlpy>=0.11.0
Requires-Dist: neo4j>=5.16.0
Requires-Dist: requests>=2.31.0
Requires-Dist: asyncio>=3.4.3
Requires-Dist: numpy>=1.24.0
Provides-Extra: dev
Requires-Dist: pytest>=7.4.0; extra == "dev"
Requires-Dist: pytest-asyncio>=0.21.0; extra == "dev"
Requires-Dist: pytest-cov>=4.1.0; extra == "dev"
Requires-Dist: black>=23.0.0; extra == "dev"
Requires-Dist: ruff>=0.1.0; extra == "dev"
Requires-Dist: mypy>=1.5.0; extra == "dev"
Requires-Dist: python-dotenv>=1.0.0; extra == "dev"
Provides-Extra: mcp
Requires-Dist: mcp[cli]>=1.0.0; extra == "mcp"
Provides-Extra: docs
Requires-Dist: mkdocs>=1.5.0; extra == "docs"
Requires-Dist: mkdocs-material>=9.0.0; extra == "docs"
Requires-Dist: mkdocstrings[python]>=0.24.0; extra == "docs"
Dynamic: license-file

# Agent Mem

A standalone Python package for hierarchical memory management in AI agents. Provides a **stateless** interface to manage active, short-term, and long-term memories with vector search, graph relationships, and intelligent consolidation.

## 🚀 Quick Links

- **📚 Documentation Index**: See [docs/INDEX.md](docs/INDEX.md) for all documentation
- **🎯 Getting Started**: See [docs/GETTING_STARTED.md](docs/GETTING_STARTED.md) for setup and testing
- **⚡ Quick Start**: See [docs/QUICKSTART.md](docs/QUICKSTART.md) for 5-minute setup
- **🐛 Bug Fixes**: See [docs/BUG_FIXES.md](docs/BUG_FIXES.md) for resolved issues
- **🏗️ Architecture**: See [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) for system design
- **👨‍💻 Development**: See [docs/DEVELOPMENT.md](docs/DEVELOPMENT.md) for implementation guide
- **📊 Progress**: See [docs/IMPLEMENTATION_STATUS.md](docs/IMPLEMENTATION_STATUS.md) for current status
- **🤖 Phase 4**: See [docs/PHASE4_COMPLETE.md](docs/PHASE4_COMPLETE.md) for AI agent integration

## ✅ Current Status

**Overall Progress**: 89% complete (98/110 major tasks completed)

**Completed Phases**:
- ✅ **Phase 1**: Core infrastructure (PostgreSQL + Neo4j, embedding service)
- ✅ **Phase 2**: Memory tiers (Active, Shortterm, Longterm repositories)
- ✅ **Phase 3**: Memory Manager (consolidation, promotion, retrieval)
- ✅ **Phase 4**: AI Agents (ER Extractor, Memory Retrieve, Memory Update)
- ✅ **Phase 5**: Testing (27 test suites - needs rewrite to match implementation)
- ✅ **Phase 9**: Streamlit UI (100% - Web UI fully functional)
- ✅ **Phase 10**: MCP Server (100% - Claude Desktop integration ready)

**In Progress**:
- 📖 **Phase 6**: Examples and demonstrations (20% complete)
- 📚 **Phase 7**: Complete API documentation (50% complete)
- 🚀 **Phase 8**: Production deployment (not started)

**See [docs/IMPLEMENTATION_STATUS.md](docs/IMPLEMENTATION_STATUS.md) for detailed progress**

## 🎨 Streamlit Web UI

**NEW**: AgentMem now includes a web-based UI for managing memories without writing code!

### Features

- 📚 **Browse Templates** - Explore 60+ pre-built BMAD templates
- ➕ **Create Memories** - Create memories using templates or custom YAML
- 📋 **View Memories** - Browse all memories for an agent
- ✏️ **Update Sections** - Edit memory sections with live Markdown preview
- 🗑️ **Delete Memories** - Safely delete with type-to-confirm protection

### Starting the UI

```bash
cd streamlit_app
streamlit run app.py
```

The UI will open at `http://localhost:8501`

### User Guide

See [docs/STREAMLIT_UI_USER_GUIDE.md](docs/STREAMLIT_UI_USER_GUIDE.md) for complete usage instructions.

---

## 🔌 MCP Server for Claude Desktop

**NEW**: AgentMem now includes a Model Context Protocol (MCP) server for integration with Claude Desktop and other MCP clients!

### Features

- 🔍 **get_active_memories** - Retrieve all active memories for an agent
- ➕ **create_active_memory** - Create new active memory with template
- 📝 **update_memory_sections** - Batch upsert multiple sections (insert/update/replace)
- 🗑️ **delete_active_memory** - Delete an active memory
- 🔎 **search_memories** - Search across memory tiers with AI-synthesized responses

### Quick Start

```powershell
# Run the MCP server (recommended: using uv)
uv run agent_reminiscence_mcp\run.py

# Alternative: using Python directly
py agent_reminiscence_mcp\run.py

# Add sample data for testing
uv run agent_reminiscence_mcp\tests\add_sample_data.py

# Test with Python client
uv run agent_reminiscence_mcp\tests\test_mcp_client.py
```

### Claude Desktop Integration

Add to your Claude Desktop config (`%APPDATA%\Claude\claude_desktop_config.json`):

```json
{
  "mcpServers": {
    "agent-reminiscence": {
      "command": "uv",
      "args": [
        "run",
        "path_to_agent_reminiscence_mcp\\run.py"
      ],
      "env": {
        "POSTGRES_HOST": "localhost",
        "POSTGRES_PORT": "5432",
        "POSTGRES_USER": "postgres",
        "POSTGRES_PASSWORD": "postgres",
        "POSTGRES_DB": "agent_reminiscence",
        "NEO4J_URI": "bolt://localhost:7687",
        "NEO4J_USER": "neo4j",
        "NEO4J_PASSWORD": "neo4jpassword",
        "OLLAMA_BASE_URL": "http://localhost:11434"
      }
    }
  }
}
```

**Important**: 
- Use `uv` command for better dependency management
- Use absolute path (adjust to your installation directory)
- Ensure environment variables match your `.env` file
- Double backslashes required in JSON on Windows

### Documentation

- **[MCP Server README](agent_reminiscence_mcp/README.md)** - Complete MCP server documentation
- **[Getting Started with MCP](docs/GETTING_STARTED_MCP.md)** - Quick start guide
- **[MCP Server Status](docs/MCP_SERVER_STATUS.md)** - Implementation status
- **[Implementation Details](docs/MCP_IMPLEMENTATION_COMPLETE.md)** - Technical details

---

## Key Features

- **Stateless Design**: One AgentMem instance can serve multiple agents/workers
- **Template-Driven Memory**: Active memories use YAML templates with structured sections
- **Section-Level Tracking**: Each section tracks its own update_count for consolidation
- **Three-Tier Memory**: Active (template+sections) → Shortterm (chunks+entities) → Longterm (temporal)
- **Simple API**: Just 4 methods to manage all your agent's memories
- **Vector Search**: Semantic search using embeddings via Ollama
- **Graph Relationships**: Entity and relationship tracking with Neo4j
- **Smart Consolidation**: Automatic section-level consolidation with conflict resolution
- **Generic ID Support**: Use any external ID (UUID, string, int) for your agents

## Quick Start

### Installation

**Linux/Mac:**
```bash
cd libs/agent_reminiscence
pip install -e .
```

**Windows:**
```powershell
cd libs\agent_reminiscence
py -m pip install -e .
```

Or add to your project's requirements:
```bash
echo "agent-reminiscence @ file:///${PWD}/libs/agent_reminiscence" >> requirements.txt
```

### Prerequisites

**Using Docker Compose (Recommended):**

The easiest way is to use the provided Docker Compose configuration:

```bash
# Start all services (PostgreSQL, Neo4j, Ollama)
docker compose up -d

# Check services are running
docker compose ps
```

**Manual Setup:**

1. **PostgreSQL** with extensions:
   - `pgvector` for vector storage
   - `pg_tokenizer` for text tokenization
   - `vchord_bm25` for BM25 search

2. **Neo4j** for graph storage

3. **Ollama** for embeddings:
   ```bash
   ollama pull nomic-embed-text
   ```

### Configuration

AgentMem supports three configuration methods:

#### Pattern 1: Direct Python (Recommended for PyPI users)

```python
from agent_reminiscence import AgentMem, Config

config = Config(
    postgres_host="localhost",
    postgres_password="secure_password",
    neo4j_uri="bolt://localhost:7687",
    neo4j_password="neo4j_password",
    ollama_base_url="http://localhost:11434"
)

agent_reminiscence = AgentMem(config=config)
```

#### Pattern 2: Environment Variables (Recommended for Docker/K8s)

```bash
export POSTGRES_HOST=postgres
export POSTGRES_PASSWORD=secure_pass
export NEO4J_URI=bolt://neo4j:7687
export NEO4J_PASSWORD=neo4j_pass
export OLLAMA_BASE_URL=http://ollama:11434
python your_app.py
```

```python
from agent_reminiscence import AgentMem

agent_reminiscence = AgentMem()  # Uses environment variables
```

#### Pattern 3: .env File (Convenient for local development)

Create a `.env` file:

```env
# PostgreSQL
POSTGRES_HOST=localhost
POSTGRES_PORT=5432
POSTGRES_USER=postgres
POSTGRES_PASSWORD=your_password
POSTGRES_DB=agent_reminiscence

# Neo4j
NEO4J_URI=bolt://localhost:7687
NEO4J_USER=neo4j
NEO4J_PASSWORD=your_password
NEO4J_DATABASE=neo4j

# Ollama
OLLAMA_BASE_URL=http://localhost:11434

# Model Configuration
EMBEDDING_MODEL=nomic-embed-text
VECTOR_DIMENSION=768

# Agent Models (Optional - defaults to Gemini)
MEMORY_UPDATE_AGENT_MODEL=google:gemini-2.0-flash
MEMORIZER_AGENT_MODEL=google:gemini-2.0-flash
MEMORY_RETRIEVE_AGENT_MODEL=google:gemini-2.0-flash
```

```python
# Automatically loaded if python-dotenv is available
from agent_reminiscence import AgentMem

agent_reminiscence = AgentMem()  # Uses .env file
```

**Note**: `python-dotenv` is optional. Install with `pip install agent-reminiscence[dev]` for .env file support.

### Basic Usage

```python
from agent_reminiscence import AgentMem
import asyncio

# Template defining memory structure
TASK_TEMPLATE = """
template:
  id: "task_memory_v1"
  name: "Task Memory"
sections:
  - id: "current_task"
    title: "Current Task"
  - id: "progress"
    title: "Progress"
"""

async def main():
    # Initialize STATELESS memory manager (serves multiple agents)
    agent_reminiscence = AgentMem()
    
    # Initialize database connections
    await agent_reminiscence.initialize()
    
    try:
        # 1. Create an active memory with template
        memory = await agent_reminiscence.create_active_memory(
            external_id="agent-123",  # Pass agent ID to method
            title="Build Dashboard",
            template_content=TASK_TEMPLATE,  # Can be YAML string or dict
            initial_sections={
                "current_task": {
                    "content": "Implement real-time analytics",
                    "update_count": 0,
                    "awake_update_count": 0,
                    "last_updated": None
                },
                "progress": {
                    "content": "- Designed UI\n- Set up project",
                    "update_count": 0,
                    "awake_update_count": 0,
                    "last_updated": None
                }
            },
            metadata={"priority": "high"}
        )
        print(f"Created memory: {memory.id}")
        
        # 2. Get all active memories for agent
        all_memories = await agent_reminiscence.get_active_memories(
            external_id="agent-123"
        )
        print(f"Total memories: {len(all_memories)}")
        
        # 3. Update multiple sections with upsert (batch operation)
        await agent_reminiscence.update_active_memory_sections(
            external_id="agent-123",
            memory_id=memory.id,
            sections=[
                {
                    "section_id": "progress",
                    "new_content": "- Designed UI\n- Set up project\n- Implemented analytics",
                    "action": "replace"  # Replace entire content
                },
                {
                    "section_id": "blockers",  # New section (upsert)
                    "new_content": "# Blockers\n- None currently",
                    "action": "replace"
                }
            ]
        )
            new_content="- Designed UI\n- Set up project\n- Implemented charts"
        )
        
        # 4. Retrieve memories (searches across all tiers)
        results = await agent_reminiscence.retrieve_memories(
            external_id="agent-123",
            query="What is the current progress on the dashboard?",
            limit=10,
            synthesis=True  # Request AI summary of findings
        )
        
        # Access results
        print(f"Mode: {results.mode}")  # "pointer" or "synthesis"
        print(f"Found {len(results.chunks)} chunks, {len(results.entities)} entities")
        if results.synthesis:
            print(f"AI Summary: {results.synthesis}")
        
    finally:
        # Clean up connections
        await agent_reminiscence.close()

if __name__ == "__main__":
    asyncio.run(main())
```

## Architecture

```
agent_reminiscence/
├── __init__.py              # Public API exports
├── core.py                  # AgentMem main class (STATELESS)
├── config/
│   ├── __init__.py
│   └── settings.py          # Configuration management
├── database/
│   ├── __init__.py
│   ├── postgres_manager.py  # PostgreSQL connection pool
│   ├── neo4j_manager.py     # Neo4j connection manager
│   ├── repositories/
│   │   ├── __init__.py
│   │   ├── active_memory.py      # Template + section CRUD
│   │   ├── shortterm_memory.py   # Shortterm memory CRUD
│   │   └── longterm_memory.py    # Longterm memory CRUD
│   └── models.py            # Pydantic models for data
├── services/
│   ├── __init__.py
│   ├── embedding.py         # Ollama embedding service
│   └── memory_manager.py    # Core memory operations (stateless)
├── agents/
│   ├── __init__.py
│   ├── memory_updater.py    # Memory Update Agent
│   ├── memorizer.py         # Memory Consolidation Agent
│   └── memory_retriever.py  # Memory Retrieve Agent
├── utils/
│   ├── __init__.py
│   └── helpers.py           # Utility functions
└── sql/
    ├── schema.sql           # PostgreSQL schema
    └── migrations/          # Future migrations
```

## API Reference

### AgentMem

The main class for memory management (STATELESS - serves multiple agents).

#### `__init__(config: Optional[Config] = None)`

Initialize stateless memory manager.

**Parameters:**
- `config`: Optional configuration object (uses environment variables by default)

#### `async initialize() -> None`

Initialize database connections and ensure schema exists.

#### `async create_active_memory(external_id: str | UUID | int, title: str, template_content: str | dict, initial_sections: Optional[dict] = None, metadata: Optional[dict] = None) -> ActiveMemory`

Create a new active memory with template-driven structure.

**Parameters:**
- `external_id`: Unique identifier for the agent (UUID, string, or int)
- `title`: Memory title
- `template_content`: Template defining section structure. Can be:
  - **Dict (JSON)**: `{"template": {...}, "sections": [{"id": "...", "description": "..."}]}`
  - **Str (YAML)**: Parsed to dict automatically (backward compatible)
- `initial_sections`: Optional initial sections that override template defaults
  - Format: `{"section_id": {"content": "...", "update_count": 0, "awake_update_count": 0, "last_updated": None}}`
- `metadata`: Optional metadata dictionary

**Returns:** Created `ActiveMemory` object with sections

**Section Structure:**
Each section contains:
- `content`: Markdown content
- `update_count`: Updates since last consolidation (resets to 0 after consolidation)
- `awake_update_count`: Total updates (never resets, for future sleep/wake features)
- `last_updated`: ISO timestamp of last update

#### `async get_active_memories(external_id: str | UUID | int) -> List[ActiveMemory]`

Get all active memories for a specific agent.

**Parameters:**
- `external_id`: Unique identifier for the agent

**Returns:** List of `ActiveMemory` objects

#### `async update_active_memory_section(external_id: str | UUID | int, memory_id: int, section_id: str, new_content: str) -> ActiveMemory`

Update a specific section in an active memory.

Automatically increments the section's `update_count` and `awake_update_count`, and triggers consolidation when threshold is reached.

**Parameters:**
- `external_id`: Unique identifier for the agent
- `memory_id`: ID of memory to update
- `section_id`: Section ID to update (defined in template)
- `new_content`: New content for the section

**Returns:** Updated `ActiveMemory` object

#### `async update_active_memory_sections(external_id: str | UUID | int, memory_id: int, sections: List[dict]) -> ActiveMemory`

Upsert multiple sections in an active memory (batch operation).

Supports creating new sections, updating existing ones, content replacement with pattern matching, and content insertion/appending.

**Parameters:**
- `external_id`: Unique identifier for the agent
- `memory_id`: ID of memory to update
- `sections`: List of section updates with structure:
  ```python
  [
    {
      "section_id": "progress",
      "old_content": "# Old",  # Optional: pattern to find
      "new_content": "# New",
      "action": "replace"  # "replace" or "insert"
    }
  ]
  ```

**Action Behaviors:**
- **replace** + no old_content: Replace entire section
- **replace** + old_content: Replace that substring
- **insert** + no old_content: Append at end
- **insert** + old_content: Insert after pattern

**Returns:** Updated `ActiveMemory` object

#### `async retrieve_memories(external_id: str | UUID | int, query: str, limit: int = 10, synthesis: bool = False) -> RetrievalResult`

Search and retrieve relevant memories for a specific agent across all tiers.

**Parameters:**
- `external_id`: Unique identifier for the agent
- `query`: Natural language search query describing the context and what information is needed
- `limit`: Maximum results per tier (default: 10)
- `synthesis`: Force AI to generate a synthesized summary of results (default: False, AI decides)

**Returns:** `RetrievalResult` with matched chunks, entities, relationships, and optional AI synthesis

#### `async close() -> None`

Close all database connections.

## Memory Tiers

### Active Memory
- **Purpose**: Template-driven working memory with sections
- **Storage**: PostgreSQL with JSONB sections
- **Structure**: JSON/YAML template + sections
  - Each section: `{content, update_count, awake_update_count, last_updated}`
- **Lifetime**: Short (hours to days)
- **Updates**: Frequent (per-section tracking with dual counters)
- **Use Case**: Current task context, ongoing work, structured progress tracking
- **Consolidation**: Section-level triggers based on update_count threshold

### Shortterm Memory
- **Purpose**: Searchable recent knowledge
- **Storage**: PostgreSQL (vectors + BM25) + Neo4j (entities/relationships)
- **Lifetime**: Medium (days to weeks)
- **Updates**: Occasional (from active memory consolidation)
- **Use Case**: Recent implementations, conversations, research findings

### Longterm Memory
- **Purpose**: Consolidated knowledge base
- **Storage**: PostgreSQL (vectors + BM25) + Neo4j (entities/relationships)
- **Lifetime**: Long (persistent)
- **Updates**: Rare (promoted from shortterm with intelligent merging)
- **Use Case**: Core knowledge, patterns, historical decisions
- **Promotion**: Automatic with type merging, confidence recalculation, and state history tracking

## Advanced Usage

### Custom Configuration

```python
from agent_reminiscence import AgentMem, Config

config = Config(
    postgres_host="custom-host",
    postgres_port=5433,
    embedding_model="custom-model",
    vector_dimension=1024,
)

agent_reminiscence = AgentMem(external_id="agent-456", config=config)
```

### Automatic Consolidation

Memory consolidation happens automatically based on update thresholds. You can also trigger it manually:

```python
# This is done internally, but exposed for advanced use cases
from agent_reminiscence.services import MemoryManager

manager = MemoryManager(external_id="agent-123")
await manager.consolidate_to_shortterm(active_memory_id=1)
await manager.promote_to_longterm(shortterm_memory_id=5)
```

### Intelligent Promotion to Longterm

When shortterm memories are promoted to longterm, the system performs intelligent merging:

**Entity Merging:**
- Matches entities by name (case-insensitive)
- Merges types from both memories (union operation)
- Recalculates confidence using weighted average (favors higher confidence)
- Tracks complete state history in metadata

**Relationship Merging:**
- Matches relationships by source and target entity names
- Merges types from both memories (union operation)
- Recalculates confidence and strength using weighted averages
- Tracks complete state history in metadata

**State History Example:**
```python
# Each entity/relationship tracks its evolution in metadata.state_history:
{
    "state_history": [
        {
            "timestamp": "2025-10-05T10:30:00Z",
            "source": "shortterm_promotion",
            "old_types": ["Person", "Developer"],
            "new_types": ["Person", "Developer", "Team Lead"],
            "old_confidence": 0.75,
            "new_confidence": 0.82
        }
    ]
}
```

This ensures no information is lost during promotion and provides a complete audit trail.

### Working with Entities and Relationships

```python
# Entities and relationships are automatically extracted during consolidation
# You can query them through the retrieval results

result = await agent_reminiscence.retrieve_memories("authentication system")

# Access entities
for entity in result.entities:
    print(f"Entity: {entity.name} (type: {entity.type})")

# Access relationships
for rel in result.relationships:
    print(f"Relationship: {rel.from_entity} -> {rel.type} -> {rel.to_entity}")
```

## Development

### Running Tests

```bash
# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/

# Run with coverage
pytest --cov=agent_reminiscence tests/
```

### Building Documentation

```bash
cd docs/
mkdocs serve  # View locally
mkdocs build  # Build static site
```

## Contributing

Contributions are welcome! Please see [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines.

## License

MIT License - see [LICENSE](LICENSE) for details.

## Support

- **Documentation**: [Full Documentation](docs/)
- **Issues**: [GitHub Issues](https://github.com/Ganzzi/agent-reminiscence/issues)
- **Discussions**: [GitHub Discussions](https://github.com/Ganzzi/agent-reminiscence/discussions)

