AI context compaction without losing important decisions

Context compaction reduces a long working history into a smaller representation for the next model call or session. The goal is to preserve decision-relevant state, not to create a smooth narrative. A concise summary that drops an approval boundary can be less useful than a longer, structured handoff.

Preserve commitments before conversation

Imagine a task where a manager approved a draft but explicitly postponed sending it. A summary saying the manager approved the message loses the distinction. The next run may treat approval of wording as permission to send.

Write the approved action, target and scope explicitly. Keep unresolved restrictions alongside progress so they do not disappear as background detail.

Use a compaction checklist

• Original objective and any later correction that changed it.
• Decisions made, including the source of authority.
• Actions completed and evidence available for each.
• Attempts with unknown outcomes that must not be repeated blindly.
• Remaining questions and the next bounded action.

Keep links to detailed records instead of copying every detail. The compacted context should tell the agent where to retrieve evidence when a decision depends on it.

Compare before and after behavior

Build a test conversation containing a changed requirement, a denied action and a partial result. Ask the system to continue once with the full history and once with the compacted version. Compare the actions it proposes, not merely whether the summaries sound similar.

If the compacted version authorizes more than the original, the summarization process is unsafe for that task. Adjust the structure and repeat the test.

Do not compact away uncertainty

Phrases such as probably delivered or needs confirmation are operational facts about evidence quality. Preserve them. Converting an uncertain attempt into a completed step can create duplicate work or false reporting. Good compaction makes the next run smaller while keeping the boundaries and unfinished work intact.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.