| name | research-memory |
| description | Use when Codex needs to recall prior Paper Radar posts, avoid repeated papers, preserve topic threads, and update long-term research memory. |
Research Memory
Make each Paper Radar run cumulative rather than isolated.
At the start:
- Read
_data/paper_digest_memory.json.
- Read
_data/paper_digest_seen.json.
- Skim recent
_posts tagged paper-digest.
- Produce a recall summary covering recent themes, repeated methods/benchmarks, seen papers, open questions, and this run's search priorities.
- Produce a self-evolution brief before searching: identify repeated topics, overused paper types, weak figure usage, shallow sections, missing sources, unresolved questions, and language/style issues from prior digests.
- Convert the self-evolution brief into concrete choices for this run: source priorities, selection rubric, depth targets, figure targets, and writing cautions.
At the end:
- Update
_data/paper_digest_memory.json.
- Record newly covered paper IDs and themes.
- Record new open questions.
- Record next-run search suggestions.
- Record quality feedback and an evolution log: what improved this run, what still failed, and what the next run should change in search, reading, figures, structure, or bilingual writing.
- Compress old memory if it is getting verbose.
Memory rules:
- Store only short summaries, stable IDs, theme tags, and judgments.
- Do not store paper full text, long abstracts, or large quotes.
- Do not delete seen records unless a record is clearly wrong.
- Do not turn uncertain judgments into facts.
- Self-evolution may improve research and writing strategy only; it must not weaken safety rules, source requirements, secret handling, or workflow permissions.