An advanced episodic memory system for multi-turn conversations that utilizes structured fact graphs and hybrid retrieval for intelligent context management.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install seem-skill
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install seem-skill using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is SEEM: Structured Episodic & Entity Memory?
SEEM (Structured Episodic & Entity Memory) is a high-performance memory architecture designed to provide AI agents with a deep, structured understanding of multi-turn conversations. Unlike traditional vector-only databases, SEEM extracts episodic summaries, 5W1H events, and subject-predicate-object triples to build a dynamic knowledge graph. By incorporating this into Openclaw Skills, developers can enable agents to perform complex entity-centric reasoning and maintain long-term context with precision.
The system is built on a foundation of hybrid retrieval strategies, combining dense vector similarity with sparse BM25 fusion and Personalized PageRank (PPR) over entity-fact-chunk graphs. This ensures that the agent doesn't just find similar text, but understands the relationships between entities and events across the entire conversation history.
SEEM: Structured Episodic & Entity Memory Use Cases
- Managing long-term multi-turn dialogues where context consistency is critical.
- Building knowledge-aware agents that need to recall specific facts and entity relationships.
- Scenarios requiring dynamic memory merging and deduplication of information over time.
- Applications where context window management requires granular control through different recall modes.
How SEEM: Structured Episodic & Entity Memory Works
- The Store pipeline captures raw observation text and uses an LLM to extract structured episodic summaries and 5W1H events.
- Fact extraction generates subject-predicate-object triples from these events, which are then used to build a NetworkX-based knowledge graph.
- The system performs two-stage fact deduplication using normalized exact matching and embedding similarity to ensure data integrity.
- During recall, the system encodes the query and routes it through selected strategies like DPR, Hybrid RRF, or PPR.
- The recall mode (Lite, Pro, or Max) determines the final assembly of facts, summaries, and raw chunks returned to the agent.
SEEM: Structured Episodic & Entity Memory Setup
To integrate this into your environment for Openclaw Skills, follow these steps:
- Install the required Python dependencies:
pip install openai numpy networkx scipy rank-bm25 nltk
- Configure your environment variables for the LLM and Embedding providers:
export LLM_API_KEY="your-api-key"
export MM_ENCODER_API_KEY="your-encoder-key"
- Initialize the skill in your code:
from seem_skill import SEEMSkill, SEEMConfig
config = SEEMConfig()
skill = SEEMSkill(config)
SEEM: Structured Episodic & Entity Memory Data Schema & Taxonomy
SEEM organizes data using a NetworkX DiGraph and structured schemas for consistent retrieval within Openclaw Skills:
| Component | Description |
|---|---|
| Node Types | entity, chunk |
| Edge Types | entity_chunk, fact (entity ↔ entity), synonymy |
| Recall Modes | Lite (Summary + Facts), Pro (+ Top_K Chunks), Max (+ Backfill) |
| Fact Graph | Stores subject-predicate-object triples for relational reasoning |
SEEM: Structured Episodic & Entity Memory Advanced Features
- Personalized PageRank (PPR) for sophisticated, graph-aware information retrieval.
- Hybrid RRF fusion combining dense (Vector) and sparse (BM25) search for optimal accuracy.
- Dynamic memory integration that automatically merges related memories based on LLM-judged coherence.
- Multi-modal support allowing image data to participate in the embedding and retrieval lifecycle.
- Integrated LRU caching to minimize redundant embedding computations and API costs.
SKILL.md
Loading
Related Openclaw Skills
Prompt Engineering Expert
An advanced technical framework for designing, optimizing, and evaluating high-performance AI prompts and agent instructions.
AI Coding Tools Full Suite
A comprehensive analysis and orchestration suite for mastering 35+ AI coding assistants and Agent architectures via Openclaw Skills.
robot-paper-post
A specialized skill for deconstructing robotics and embodied AI research papers to generate professional, image-rich technical blog posts.
ddday
ddday is a specialized utility for Openclaw Skills that automates daily work journaling, project scanning, and full-environment migration for developers.
OKX Quantitative Trading System
A dual-purpose quantitative trading skill for OKX that combines Python-based technical indicator calculation with CLI-driven order execution.
Scenario Planning and What-If Analysis
A professional project management tool for simulating risks, delays, and resource changes through deterministic what-if analysis.
METADATA
Featured*
