nbj-ob1-agent-memory-openclaw

Use the OB1 agent memory workflow: call openbrain_recall before meaningful work with task type, query, and scope, then call openbrain_writeback after completion with compact decisions, lessons, and next steps. This lets later tasks or agents continue without rereading transcripts.

Each memory carries a use_policy: can_use_as_instruction memories guide behavior directly, can_use_as_evidence memories inform reasoning only, and requires_user_confirmation memories must be surfaced before reliance. Agent-written memory starts as evidence and becomes instruction only after human confirmation or trusted import.

Write back compact categories: decisions, outputs, lessons, constraints, unresolved questions, next steps, failures, and artifacts. Never store raw transcripts, reasoning traces, secrets, credentials, large code blocks, or private data dumps; store summaries and source references instead.

The agent continues the task normally and notes that no OB1 recall or write-back occurred. It must not invent remembered facts or fabricate prior context when the openbrain tools cannot be reached.

For pull request reviews, the agent recalls repo conventions, prior review comments, recurring bug patterns, and maintainer preferences, then writes back recurring issue patterns, false positives, and reusable repo lessons. Full diffs are never stored; only artifact references to PRs, commits, or files.