Project Amber
Room
Memory that doesn't reset.
Overview
Long-context memory architecture research aimed at making an agent's state feel continuous across long-horizon sessions rather than resetting per interaction.
Research Focus
The Research Problem
Standard RAG (Retrieval-Augmented Generation) retrieves isolated text blocks based on vector similarity, losing temporal continuity, narrative state progression, and long-term goal dependencies across multi-turn agent interactions.
Why It Matters
Autonomous agents handling long-horizon tasks (e.g. multi-week codebase refactoring or continuous system monitoring) require persistent working memory that reinforces important context while decaying obsolete state.
Research Objective
Design structured memory graphs and context compression algorithms that preserve continuous narrative state, goal tracking, and episodic decay across sessions.
Technical Approach
Key Methodology Vectors
- Graph-based memory architectures with temporal edges connecting key decision nodes.
- Reinforcement and decay weighting to prioritize active context while pruning stale data.
- Hierarchical summarization loops that maintain top-level objectives alongside granular execution details.
Expected Outcomes & Milestones
- Low-latency continuous memory middleware enabling agents to maintain operational coherence across multi-week sessions.
- Reduction in context-window overflow errors during long-horizon agent execution.
References & Prior Work
- MemGPT: Towards LLMs as Operating Systems — Packer et al., arXiv:2310.08560 (2023).
- Episodic Memory Graph Blueprints for Autonomous Agents — Sleepers Research Internal Paper (2026).
Interested in this research direction?
Sleepers Research welcomes inquiries from teams exploring autonomous agent security and applied machine learning.
Contact the Laboratory