Memory enables Agents to maintain context across interactions, store factual knowledge, and recall past events or user preferences.
Core loop extends to: Observe → Retrieve relevant memories → Think → Act → Store new memories → Repeat
Memory Types
Conversation history
Working memory (scratchpad/tool outputs),
long-term memory (external storage, often backed by vector databases).
Current State
Memory remains the key bottleneck for truly autonomous, long-running Agents .
Context windows are expanding but still too limited for lifelong learning; retrieval-augmented generation (RAG) is the standard workaround.
Simple vector similarity retrieval handles many tasks, but agents struggle with temporal reasoning, conflict resolution, and memory compression over time.
Storing detailed personal history is expensive and raises data sovereignty concerns. Local-first and user-managed stores are emerging as partial answers.