Announcing Observational Memory
Source: https://mastra.ai/blog/observational-memory Author: Mastra Date: 2026-02-10
Summary
Mastra (open-source agent framework) announces observational memory — a system where agents automatically extract and store relevant facts from conversations without explicit instruction. Different from Letta’s git-based memory: Mastra’s is conversational and automatic.
Key Claims
- Observational memory: as the agent converses, it automatically identifies and stores “facts” in a structured memory store.
- No user action required: extraction is automatic based on salience detection (what seems important to remember).
- Retrieval: at query time, relevant memories are fetched and injected into context — similar to RAG but for agent memories.
- Key design choice: store structured facts, not raw conversation history. Reduces token cost, improves precision.
- Open question: what to forget? Mastra’s approach: recency + retrieval frequency determine memory retention.
Connection to Other Sources
Different approach to agent memory than Letta (file-based) or mem-agent (RL-trained). All three are addressing the same problem: what should agents remember across sessions?
Concepts
- Agent Memory — observational/automatic extraction approach
- Context Engineering — memory injection into context