Home / Technology / AI Memory Breakthrough: Agents Recall Past Decisions
AI Memory Breakthrough: Agents Recall Past Decisions
11 Feb
Summary
- New observational memory compresses conversation history into logs.
- Stable context windows reduce token costs by up to 10x.
- System prioritizes past agent decisions over broad knowledge recall.
Modern AI workflows are moving beyond limitations of Retrieval-Augmented Generation (RAG), especially for long-running, tool-heavy agents. Observational memory, an open-source technology, offers an alternative architecture focusing on persistence and stability.
This system employs two background agents to compress conversation history into a dated observation log, eliminating dynamic retrieval. It boasts text compression of 3-6x and 10-40x for large tool outputs. This method prioritizes recall of agent decisions over broad external corpus searching.
The architecture divides the context window into observations and raw message history. The Observer agent compresses messages when they reach a token threshold, and the Reflector agent restructures the observation log. This process requires no specialized databases, unlike vector or graph databases.
Observational memory significantly cuts token costs, up to 10x, by maintaining stable context windows that allow for aggressive prompt caching. This stability is crucial for production teams facing unpredictable costs with dynamic retrieval systems.
Unlike traditional compaction methods that produce documentation-style summaries, observational memory creates an event-based decision log. This log captures specific decisions and actions, providing a more detailed and actionable history for agents.
This technology is ideal for enterprise use cases like in-app chatbots and AI SRE systems, where long-running conversations and consistent context maintenance over weeks or months are critical. Mastra 1.0 includes this technology, with plugins available for frameworks like LangChain and Vercel's AI SDK.
