Validated patterns from the community
LLM context windows are finite. Long conversations get truncated, losing important early context. Agents "forget" decisions made earlier in the session, leading to contradictions or repeated work.