Agent should auto-save expensive research to memory files before context compaction

Problem

When Claude Code performs long, token-intensive research tasks (e.g., web searches, document analysis, terminology comparison across multiple sources), the results exist only in the conversation context. When the context window fills up and gets compacted/summarized, all the detailed research is lost — only a brief summary survives.

The agent has access to a persistent memory/ directory specifically designed for cross-session knowledge retention, but it does not proactively save research results there.

Real-world example

I asked Claude to research Japanese game development terminology across Unity, RPG Maker, Wolf RPG Editor, and SRPG Studio before translating Construct 3 localization files. This took over 1 hour of web searches, document fetching, and comparative analysis, consuming a significant amount of tokens.

The research produced valuable results — a detailed terminology comparison table with sources. However:

1. Only a 6-line summary was written into the code (LANG_RULES dictionary)
2. The detailed comparison table, sources, and reasoning were not saved to memory/
3. When context was compacted, all details were lost
4. When I later asked about a specific term ("Destroy"), the agent had to redo the research from scratch, wasting more tokens

Root cause

The auto-memory guidelines say "save stable patterns confirmed across multiple interactions" — this framing is too conservative and causes the agent to under-value expensive one-time research that is equally worth preserving.

Expected behavior

When the agent performs substantial research (multiple web searches, document fetches, cross-referencing), it should:

1. Proactively save detailed findings to memory/ files — not just use them in the moment
2. Prioritize saving when research has consumed significant tokens (e.g., >5 web search/fetch calls)
3. At minimum, save before context gets large enough that compaction is likely

Suggested improvement

Add guidance in the auto-memory system prompt:

When you perform research involving multiple web searches or document analysis (>5 search/fetch calls), proactively save the structured results to a topic-specific memory file. Don't wait for the user to ask — treat expensive research output as an asset worth preserving.

Environment

• Claude Code CLI
• Model: Claude Opus 4.6
• Platform: Windows 11

🤖 Generated with Claude Code