DREAM Framework

DREAM is a memory-centric framework for AI agents, designed to provide long-term episodic memory with a modular and extensible architecture.

It is based on the DREAM Architecture, a research proposal focused on efficient memory organization, reduced energy usage, and scalable cognition for intelligent systems.

This repository provides a reference implementation of the DREAM architecture as a Python framework.

---

Why DREAM?

Modern AI agents and LLM-based systems suffer from fundamental limitations:

• Short context windows
• Loss of long-term episodic information
• High energy and memory costs when scaling context
• Tight coupling between memory, model, and infrastructure

DREAM addresses these issues by treating memory as a first-class system component, decoupled from models and execution logic.

---

Key Concepts

• Episodic Memory: Memories are stored as discrete episodes with metadata.
• Backend Abstraction: Memory storage and retrieval are fully pluggable.
• Low Coupling: Agents interact with memory through a simple, stable API.
• Research-Oriented: Designed to support experimentation and future backends (vector databases, embeddings, distributed storage).

---

Installation

For now, DREAM is intended to be installed in editable mode:

pip install -e .

PyPI distribution will be available in a future release.

---

Quick Example

from dream import Memory

memory = Memory()

memory.ingest(
    "User prefers narrative gameplay and dislikes long combat",
    {"user_id": "123"}
)

results = memory.retrieve("combat")

for r in results:
    print(r["content"])

---

Using a Custom Backend

DREAM supports backend injection by design.

from dream import Memory
from dream.backends.inmemory import InMemoryBackend

backend = InMemoryBackend()
memory = Memory(backend=backend)

Any backend implementing the BaseMemoryBackend interface can be used.

---

Architecture Overview

The framework follows a clean separation of concerns:

• Memory: Public API used by agents
• BaseMemoryBackend: Abstract backend interface
• InMemoryBackend: Reference backend implementation
• examples/: Usage demonstrations
• tests/: Minimal validation tests

This design allows new backends (SQLite, FAISS, vector databases, etc.) to be introduced without modifying agent code.

---

Project Status

• Current version: v0.1.0
• Status: Early-stage framework / reference implementation
• Focus: Correct architecture, extensibility, and clarity

This project is intentionally minimal at this stage. Performance optimizations and advanced backends will be introduced incrementally.

---

Relation to the DREAM Architecture

This framework is a practical implementation inspired by the DREAM research work:

• DREAM Architecture (Paper): https://zenodo.org/records/17619917

The goal of this repository is not to replace the paper, but to provide a concrete and extensible foundation for experimentation and further research.

---

Roadmap (High-Level)

• Core memory API [✓]
• Backend abstraction [✓]
• In-memory backend [✓]
• Persistent storage backend (SQLite)
• Vector-based retrieval backend
• Embedding-based similarity
• CLI utilities
• PyPI release

---

License

This project is licensed under the Apache License 2.0.

---

Author

Matheus Pereira da Silva Independent Researcher & Software Developer

---

Disclaimer

DREAM is a research-oriented framework. It is not intended to be a production-ready system at this stage. Use it for experimentation, learning, and exploration.