Claude Curator Plugin

A knowledge curation agent for Claude Code with memory retrieval that surfaces past learnings based on current context.

Acknowledgments

This plugin was inspired by claude-diary by Lance Martin, which demonstrates long-term memory management through diary entries and reflection. The curator plugin adapts similar concepts for research curation with intent-based triggering and cross-referenced knowledge building.

Overview

The Curator plugin helps you maintain organized, searchable research logs across your projects. Unlike simple note-taking, it builds a knowledge graph through cross-references and always consults past research before conducting new investigations.

Key Features

• Memory Retrieval: Automatically surfaces relevant past research before new investigations
• Hooks: Optional hooks for passive memory retrieval and session summaries
• Cross-Referencing: Links related entries with Related and Builds On fields
• Structured Knowledge: Maintains research.md with Active Research, Completed Research, Research Log, Ideas, and References
• Tag-Based Organization: Consistent tagging for easy retrieval
• Gap Detection: Identifies what's known vs what needs new research

Architecture

curator/
├── .claude-plugin/
│   └── plugin.json          # Plugin metadata
├── agents/
│   └── curator.md           # Curator agent with memory retrieval
├── commands/
│   └── curate.md            # /curate slash command
├── hooks/
│   ├── pre-compact.sh       # Session summary before compaction
│   ├── memory-retrieval.py  # Intent-based memory retrieval
│   └── settings.json        # Example hook configuration
├── examples/
│   └── sample-research.md   # Example research.md structure
├── README.md
├── INSTALL.md
└── LICENSE

How It Works

Memory Retrieval Flow

User: /curate agent coordination patterns

                    ↓

┌─────────────────────────────────────────────────────────┐
│ 1. MEMORY RETRIEVAL (Always First)                      │
│    - Extract concepts: [agent, coordination, patterns]  │
│    - Search research.md by tags and keywords            │
│    - Rank by: tag match > keyword > recency             │
│    - Surface top 3-5 related entries                    │
└─────────────────────────────────────────────────────────┘
                    ↓
┌─────────────────────────────────────────────────────────┐
│ 2. GAP IDENTIFICATION                                   │
│    - What's already known from past research?           │
│    - What knowledge gaps remain?                        │
│    - Present to user for confirmation                   │
└─────────────────────────────────────────────────────────┘
                    ↓
┌─────────────────────────────────────────────────────────┐
│ 3. FOCUSED RESEARCH                                     │
│    - WebSearch targeting identified gaps                │
│    - WebFetch for detailed extraction                   │
│    - Synthesize with awareness of past findings         │
└─────────────────────────────────────────────────────────┘
                    ↓
┌─────────────────────────────────────────────────────────┐
│ 4. CROSS-REFERENCED UPDATE                              │
│    - Add new entry with Related links                   │
│    - Include Builds On references                       │
│    - Update past entries with forward links             │
│    - Maintain tag index                                 │
└─────────────────────────────────────────────────────────┘

Example Session

User: /curate agent coordination patterns

Curator:
I found related research in your knowledge base:

1. **Context Engineering for Multi-Agent Systems** (2025-12-15) - Active Research
   Key insight: Context passing can replace explicit coordination in some cases
   Tags: #agent #context #architecture

2. **Stateless Agent Design** (2025-12-10) - Completed Research
   Key insight: Stateless agents with shared context eliminate coordination overhead
   Tags: #agent #architecture #stateless

Based on this, I'll focus new research on:
- Explicit coordination mechanisms (not covered by past research)
- Cases where context-passing isn't sufficient
- Hybrid approaches

[Conducts focused research...]

Added to Active Research:
- **Explicit Agent Coordination Mechanisms** (2026-01-02)
- Related: Context Engineering, Stateless Agents
- Builds On: Past work showed context-passing works for shared state;
  this covers cases where explicit coordination is still needed

Usage

Basic Commands

When loaded via --plugin-dir, commands are namespaced as /curator:curate:

# Research a topic (with memory retrieval)
/curator:curate agent coordination patterns

# Show current research status
/curator:curate

# Quick note (auto-links to related research)
/curator:curate note: context engineering eliminates coordination in some cases

# Save a reference
/curator:curate ref: https://example.com/paper - great overview of patterns

# Capture an idea for later
/curator:curate idea: explore graph-based knowledge representation

# Update existing research
/curator:curate update: mark context engineering as completed

# Pure memory retrieval (no new research)
/curator:curate recall: what do we know about stateless agents

# Extend past research
/curator:curate extend: stateless agents - what about hybrid approaches?

Note: If installed globally to ~/.claude/plugins/, the command may be available as just /curate.

Natural Language Triggers

The curator activates automatically for:

• "Research [topic]"
• "Find information about [topic]"
• "What's the latest on [topic]?"
• "What do we know about [topic]?"
• "Build on [past topic]"
• "Document this idea"

research.md Structure

Each project maintains a research.md with:

# Research Log

> Project: my-project
> Started: 2025-12-01
> Last Updated: 2026-01-02

## Active Research
[Ongoing investigations with status, findings, next steps]

## Completed Research
[Archived research with conclusions and key insights]

## Research Log
[Chronological record of all activities - newest first]

## Ideas & Future Research
[Parking lot for future exploration]

## References Library
[Categorized references with By Category and By Tag indices]

Cross-Reference Fields

New in this version:

### 2026-01-02 10:30 - Agent Coordination Patterns
**Tags**: #agent #coordination #patterns
**Type**: Research
**Related**: [Context Engineering](#2025-12-15), [Stateless Agents](#2025-12-10)
**Builds On**: Previous research on context-passing; extends to explicit coordination cases

[Content...]

Hooks (Optional)

The plugin includes two hooks for automatic memory retrieval without explicit commands:

PreCompact Hook

Trigger: Before conversation compaction (200+ messages or manual /compact)

Behavior: Prompts Claude to add a session summary to research.md, preserving:

• Topics researched this session
• Key insights discovered
• References added
• Open questions to continue next session

UserPromptSubmit Hook

Trigger: Only prompts with curation intent (not every prompt)

Intent patterns that trigger the hook:

Category	Patterns
Research	"research X", "find information about", "look up", "what's the latest", "curate", "what do we know"
Ideas	"idea:", "brainstorm", "I'm thinking about", "what if we", "document this idea"
References	"save this", "ref:", any URL, "add this to research", "track this"
Updates	"update:", "note:", "mark completed", "update research"

Non-triggers (examples):

• "Fix the bug in main.py" - coding task
• "Run the tests" - command
• "What does this function do?" - code question

Behavior when triggered:

1. Extracts keywords from your prompt
2. Searches research.md for matching entries (by tags and content)
3. Injects top 3 matches as context: [Curator] Related past research found: ...

Example

You: "Research agent coordination patterns"

[Hook detects "research" intent, searches research.md, finds matches]

Claude sees: "[Curator] Related past research found: Agent Coordination Patterns,
Context Engineering for Multi-Agent Systems, Stateless Agent Design"

Claude: "Based on your past research on agent coordination..."

You: "Fix the authentication bug"

[Hook detects NO curation intent, exits silently]

Claude: Proceeds with bug fix without curator context

See INSTALL.md for hook installation instructions.

---

Technical Details

Memory Retrieval Algorithm

1. Concept Extraction: Parse query into key terms and related concepts
2. Tag Search: Grep for exact and related tag matches
3. Keyword Search: Search titles and content
4. Relevance Ranking:
   • Exact tag match: highest weight
   • Related tag match: high weight
   • Keyword in title: medium weight
   • Keyword in content: lower weight
   • Recency bonus for fast-moving topics
5. Threshold: Surface entries with relevance score above threshold

Tag Taxonomy

Universal: #research, #reference, #idea, #insight, #note

Domain: #architecture, #security, #performance, #api, #ml, #agents, #context, #devops, #testing, #ux

Technology: #python, #typescript, #claude, etc.

Status: #active, #completed, #blocked, #validated

Comparison: Before vs After

Aspect	Before (Basic Curator)	After (With Memory Retrieval)
Research Start	Fresh search every time	Search past research first
Knowledge Building	Isolated entries	Cross-referenced graph
Duplicate Prevention	Manual	Automatic gap detection
Context Awareness	None	Surfaces relevant past work
Entry Structure	Basic	Related + Builds On fields

Installation

See INSTALL.md for detailed installation instructions.

Quick Start (Local Testing)

claude --plugin-dir /path/to/curator

Global Installation

cd ~/.claude/plugins
git clone https://github.com/sesh11/curator

Limitations

• Single Project Scope: research.md is per-project by design
• Tag Fragmentation: No automatic consolidation of similar tags (#ml vs #machine-learning)
• Flat Structure: Markdown-based, not a true knowledge graph
• Manual Archiving: Large files require manual archive decision

Future Improvements

• Cross-Project Memory: Optional global research index
• Tag Normalization: Automatic consolidation of similar tags
• Confidence Scoring: Weight sources by credibility
• Contradiction Detection: Flag conflicting findings
• Knowledge Graph Backend: Graph structure rendered as markdown

License

MIT