raptor-memory

name	raptor-memory
description	Replace OpenClaw's flat-file memory with a self-learning knowledge engine featuring RAPTOR trees, knowledge graphs, conflict resolution, and importance scoring. Your agent actually gets smarter over time.
homepage	https://github.com/incidentfox/self-learning-ai-agent

RAPTOR Memory — Self-Learning Knowledge Engine

You have access to an intelligent memory system that goes far beyond storing text in files. Use these MCP tools instead of writing to memory files whenever you need to remember, recall, or reason about past information.

When to use these tools

Learning new information

When you encounter important information (from conversations, files, meetings, research), use learn_knowledge to store it. The engine will:

• Extract structured knowledge (facts, procedures, relationships, decisions)
• Detect duplicates and conflicts with existing knowledge
• Auto-resolve or flag contradictions
• Build hierarchical summaries (RAPTOR tree) for multi-level abstraction

Answering questions from memory

When the user asks about something you should know from past interactions, use query_knowledge. The engine will:

• Search across semantic similarity, knowledge graph relationships, and hierarchical summaries
• Return grounded answers with source citations and confidence scores
• Synthesize across multiple knowledge sources (not just top-1 match)

Teaching explicit facts

When the user explicitly tells you to remember something, use teach_knowledge. This creates high-confidence knowledge nodes that persist and are prioritized in retrieval.

Searching for specific topics

Use search_knowledge when you need to find specific knowledge nodes (e.g., to verify what you know about a topic before answering).

Detecting corrections

When someone says "actually that's wrong" or corrects previous information, use detect_correction to identify and learn from the correction.

Why this is better than file-based memory

Capability	File Memory	RAPTOR Memory
Storage	Markdown files	Structured knowledge graph + vector embeddings
Search	BM25 + basic vector	Multi-strategy (semantic + graph + hierarchical)
Conflicts	Last write wins	5x5 conflict resolution matrix
Abstraction	None	RAPTOR tree (leaf → summary → overview)
Importance	Manual curation	Observation-driven scoring with temporal decay
Learning	Append-only	Continuous learning with dedup + merge

Important guidelines

• Always use learn_knowledge for information worth remembering long-term
• Use query_knowledge before answering questions about past events or decisions
• When you detect a correction in conversation, call detect_correction
• Check knowledge_stats periodically to understand what you know
• Trust the confidence scores — low confidence means you should caveat your answer