OpenHands Persistent Memory Flow Diagram

Who writes, who reads

No memory tool or store: the agent edits Markdown files with its normal tools, guided by the <MEMORY> prompt section
LocalConversation reads both indexes once, when the agent is first made ready (best-effort: read errors are ignored)
Daily logs are never injected; MEMORY.md may point to them and the agent reads them on demand

Staleness and trust

Over budget: whole lines are dropped from the top, so the newest appended facts survive
Content is wrapped in <UNTRUSTED_CONTENT> and labelled as possibly stale hints
memory_context is excluded from serialization, so it is re-read from disk every session

Default without load_memory

load_memory defaults to False; the <MEMORY> section then tells the agent to use repository AGENTS.md as memory
AGENTS.md reaches the prompt as an always-on repo skill, not through load_memory