菜单
Defragment and clean research agent memory - consolidate paper info, organize research domains, remove redundancy from research tracking.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/acertainKnight/project-thoth --skill research-memory-cleanup命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | research-memory-cleanup |
| description | Defragment and clean research agent memory - consolidate paper info, organize research domains, remove redundancy from research tracking. |
Clean up research agent memory by consolidating duplicate paper info, organizing research domains, and removing redundancy.
Most common use: Memory has grown messy with duplicate paper entries, scattered research notes, or conflicting discovery configs.
User: "My research memory is getting messy."
Agent: "Let's clean it up! I'll:
1. Backup your current memory (safety first)
2. Review what needs consolidating
3. Clean up duplicate/scattered info
4. Restore the cleaned version
This takes ~5 minutes. Ready to start?"
Then execute the 3-step workflow.
Ask 3 questions to understand the mess:
Question 1: What's messy?
Agent: "What feels messy in your research memory?
A) Duplicate paper entries
B) Scattered research domain notes
C) Conflicting discovery configs
D) Old/outdated research areas
E) General disorganization"
Question 2: How bad is it?
Agent: "Scale of messiness:
A) Mildly cluttered (50-100 papers tracked)
B) Pretty messy (100-300 papers)
C) Very messy (300+ papers or 50K+ chars)
D) Complete chaos (multiple research agents merged)"
Question 3: What to preserve?
Agent: "What must NOT be lost?
- Active research questions?
- Discovery configurations?
- Paper reading lists?
- Citation networks?
- Analysis notes?
Tell me what's sacred, I'll protect it."
Trigger: General memory mess, not critical
Agent: "Starting standard cleanup. This will:
Step 1: Backup everything
Step 2: Consolidate duplicate paper entries
Step 3: Organize research domains
Step 4: Clean discovery configs
Step 5: Restore cleaned version
Backing up now..."
[Run backup script]
Agent: "Backup saved to .letta/backups/[agent-id]/[timestamp]/
Now spawning memory subagent to clean up the files..."
[Spawn memory subagent with cleanup task]
Agent: "Memory subagent working... (this takes 2-3 min)
[Wait for subagent to finish]
Agent: "Cleanup complete! Changes:
- Merged 15 duplicate paper entries → 5 unique papers
- Organized 3 research domains (ML, NLP, RL)
- Removed 2 outdated discovery configs
- Total reduction: 12,450 chars → 8,200 chars (34% smaller)
Restoring cleaned memory now..."
[Run restore script]
Agent: "Done! Your research memory is now organized and deduplicated."
Trigger: Many duplicate paper entries
Agent: "I see lots of duplicate paper entries. Cleaning up papers...
[Memory subagent task: consolidate papers]
Consolidation strategy:
1. Group by paper title/DOI
2. Merge all mentions into single entry
3. Keep: title, authors, citations, your notes
4. Remove: duplicate tracking, scattered mentions
Subagent will merge papers and DELETE source duplicates."
Memory subagent prompt:
Edit memory files in .letta/backups/working/ to consolidate paper entries.
Find duplicate papers (same title/DOI):
1. Create single comprehensive entry per paper
2. Merge all notes, citations, analysis
3. DELETE the duplicate source files (use rm command)
Example:
- paper-transformers-1.md
- paper-transformers-duplicate.md
→ Merge into single paper-transformers.md
→ rm paper-transformers-duplicate.md
After consolidation, report what was merged.
Trigger: Research areas scattered, no clear structure
Agent: "Your research domains are scattered. Let me organize them...
[Memory subagent task: organize domains]
Organization strategy:
1. Identify main research areas (ML, NLP, CV, etc.)
2. Group papers/notes by domain
3. Create domain-specific memory blocks
4. Link related domains
Subagent will create organized structure."
Memory subagent prompt:
Organize research memory by domain.
Create domain blocks:
- research-domain-ml.md (ML papers and notes)
- research-domain-nlp.md (NLP papers and notes)
- research-domain-rl.md (RL papers and notes)
Move relevant content into each domain block.
DELETE scattered source files after merging.
Use clear structure:
## Papers
- Paper 1: [title] - [notes]
- Paper 2: [title] - [notes]
## Key Concepts
- Concept 1: [description]
## Active Questions
- Question 1: [question]
Trigger: Old/conflicting discovery configurations
Agent: "I see multiple discovery configs, some outdated. Cleaning...
[Memory subagent task: clean discovery configs]
Cleanup strategy:
1. Keep: Current active discovery configs
2. Archive: Old configs (if might need later)
3. Remove: Clearly outdated/conflicting configs
Subagent will preserve active, remove dead configs."
Every cleanup follows this:
# Step 1: Backup
npx tsx [SKILL_DIR]/scripts/backup-memory.ts $LETTA_AGENT_ID .letta/backups/working
# Step 2: Clean (spawn memory subagent with task)
Task({
subagent_type: "memory",
description: "Clean research memory",
prompt: "[specific cleanup task]"
})
# Step 3: Restore
npx tsx [SKILL_DIR]/scripts/restore-memory.ts $LETTA_AGENT_ID .letta/backups/working
Never skip backup!
Common research memory issues:
| Issue | Consolidation Strategy |
|---|---|
| Duplicate papers | Merge by DOI/title → single entry |
| Scattered domains | Group by topic → domain blocks |
| Old discovery configs | Keep active, archive old, delete dead |
| Redundant notes | Merge similar notes → single coherent note |
| Conflicting info | Resolve conflicts → choose correct version |
| Walls of text | Add structure → headers + bullets |
Template for cleanup tasks:
Edit memory files in .letta/backups/working/ to [specific goal].
Focus on:
- [Primary consolidation task]
- [Secondary organization task]
- [Tertiary cleanup task]
IMPORTANT: After merging content, DELETE source files (rm command).
You have full bash access in .letta/backups/working/.
Delete files when:
1. Content merged into another block
2. File is duplicate/junk data
3. File is outdated and irrelevant
Report what was merged and what was deleted.
Key principles:
If cleanup goes wrong:
# Find the backup
ls -la .letta/backups/[agent-id]/
# Restore from timestamp
npx tsx [SKILL_DIR]/scripts/restore-memory.ts $LETTA_AGENT_ID .letta/backups/[agent-id]/[timestamp]
Pro tip: Each backup is timestamped, so you can roll back to any previous state.
Symptom: Important research notes were removed
Fix: Roll back and be more specific
Agent: "The cleanup removed too much. Let me roll back:
[Restore from backup]
Agent: "Rolled back to pre-cleanup state.
Let's try again with more specific instructions:
- What exactly should be consolidated?
- What must be preserved no matter what?
Tell me the specifics and I'll give clearer instructions to the memory subagent."
Symptom: Cleanup didn't merge all duplicates
Fix: More specific merge instructions
Agent: "Some papers are still duplicated. Let me:
1. Identify the duplicates (same DOI/title)
2. Show you which ones
3. Give memory subagent explicit merge list
This will be more precise."
Enhanced prompt:
Merge these specific paper duplicates:
- paper-bert.md + paper-bert-2.md → paper-bert.md
- paper-gpt3.md + gpt3-analysis.md → paper-gpt3.md
For each merge:
1. Combine all unique info
2. DELETE the duplicate file
3. Report: "Merged [file1] + [file2] → [result]"
Symptom: Cleanup created walls of text, no organization
Fix: Add structure requirements
Agent: "The cleanup lost structure. Let me fix that:
[Re-run with structure requirements]
Memory subagent will now:
- Use markdown headers (##, ###)
- Organize with bullet points
- Keep related info together
- Add scannable formatting
Running again..."
Symptom: One cleanup pass isn't enough
Fix: Multiple passes with different focuses
Agent: "This needs multiple cleanup passes. Let's do:
Pass 1: Consolidate papers (done)
Pass 2: Organize domains (next)
Pass 3: Clean configs (after)
Each pass builds on the previous. Ready for Pass 2?"
| Problem | Quick Fix |
|---|---|
| Backup failed | Check disk space, .letta/backups/ permissions |
| Subagent deleted too much | Rollback, be more specific about what to preserve |
| Restore failed | Check working directory has cleaned files |
| Memory still messy | Run multiple passes, each with specific focus |
| Lost important notes | Rollback immediately, check backup timestamp |
| Cleanup took forever | Reduce scope, clean one section at a time |
| Files not deleted | Add "use rm command" explicitly to prompt |
| No consolidation | Memory subagent needs clearer merge instructions |
| Structure lost | Add formatting requirements to prompt |
Use case: Reading list has 50+ papers, many duplicates
Agent: "Cleaning up paper reading list...
Strategy:
1. Group by topic/domain
2. Mark status (read, reading, to-read)
3. Remove duplicates
4. Sort by priority
Memory subagent will create organized reading-list.md."
Use case: Citation graphs scattered across files
Agent: "Consolidating citation networks...
Strategy:
1. Merge all citation data
2. Remove duplicate edges
3. Update paper metadata
4. Create single citation-graph.md
Memory subagent will consolidate graphs."
Use case: Discovery results from 6 months, many outdated
Agent: "Cleaning discovery history...
Strategy:
1. Keep: Last 30 days of discoveries
2. Archive: 30-90 days (compress to summaries)
3. Delete: 90+ days (unless marked important)
Memory subagent will prune old discoveries."
Core workflow:
Key principles:
What to consolidate:
What to preserve:
Success: Research memory is organized, deduplicated, and structured - agent can quickly access papers, domains, and configs without wading through duplicates.
Initialize a new user's research assistant. Use this on first interaction or when user asks to "get started", "set up", or "introduce yourself". Also use when you don't know the user's research interests or the human memory block still has placeholder text.
Walk the user through new Thoth features since their last onboarding or update. Use when the user asks "what's new", "what changed", or "what can you do now". Also use after check_whats_new returns updates to walk through them.
Create, manage, and iterate on research plan documents in the Obsidian vault. Use when the user asks for a research plan, literature review roadmap, or when you need to formalize your own working research strategy.