Observer-Reflector Architecture
Memory system architecture used in Mastra's Observational Memory with two background agents that compress and garbage collect conversation history achieving 5-40x compression.
About this tool
Overview
The Observer-Reflector Architecture is a two-agent system used in Mastra's Observational Memory that watches conversations and maintains a compressed text-only observation log.
Architecture Components
Observer Agent
When messages hit 30k tokens (configurable threshold), a separate "observer agent" compresses messages into new observations that are appended to the observation block. The compression is typically 5–40×.
Reflector Agent
When observations hit 40k tokens (also configurable), a separate "reflector agent" garbage collects observations that don't matter, removing outdated or irrelevant information.
Context Window Structure
The context window is broken into two blocks:
- First block: List of compressed observations
- Second block: Raw messages that haven't yet been compressed
Key Benefits
- Completely stable context window that's predictable and reproducible
- Prompt-cacheable across many agent/user turns
- Uses formatted text rather than structured objects
- Easier to debug than vector-based approaches
Performance
Observational Memory achieves 94.87% on LongMemEval with GPT-5-mini, the highest score ever recorded on this benchmark.
Pricing
Open-source as part of Mastra framework.
Loading more......
Information
Categories
Tags
Similar Products
6 result(s)