An agent transcript is a flight recorder. It preserves prompts, responses, tool calls, errors, output, branches, and abandoned paths. That history is valuable evidence, but retaining every event does not make it useful memory.

Memory requires judgment. A megabyte of tool output may matter for replay and be worthless tomorrow, while one sentence explaining why an architecture was rejected may remain valuable for months. A memory layer should extract, scope, retrieve, supersede, and forget state instead of treating the session log as an ever-growing source of truth.

The useful pipeline is evidence to judgment to durable state to retrieval. This keeps session history available for audit while allowing the next agent to receive a small, relevant, maintainable representation of what actually matters.