generic-llm-memorizer 0.1.5

A library for managing LLM conversations and context

Generic LLM Memorizer

Library for managing AI agent memory through automatic fact extraction, knowledge graph building, and SKOS ontology storage.

Features

• Fact Extraction: Automatically extracts relevant facts from conversations (preferences, skills, context, events)
• User Profile: Builds and maintains a user profile (name, age, hobbies, occupation)
• Conversation Summary: Generates concise summaries of conversations
• Knowledge Graph: Builds an RDF-style knowledge graph as triplets (Subject → Predicate → Object)
• SKOS Knowledge Base: Stores ontological concepts following the SKOS standard
• Search: Semantic search through facts and SKOS concepts
• Persistence: All data stored in JSON format

Installation

uv sync

Usage

Basic Setup

import asyncio
from generic_llm_memorizer.llm_memorizer import LLMMemorizer
from pydantic import BaseModel

# Define your LLM call function
async def llm_call(prompt: str, system_prompt: str, response_model: type[BaseModel]):
    # Implement your LLM call here (OpenAI, Anthropic, Ollama, etc.)
    # Return an instance of response_model
    pass

# Initialize memorizer with a list of conversation messages (strings)
memorizer = LLMMemorizer(
    user_id="user_123",
    session_id="session_1",
    llm_call=llm_call,
    conversation=[
        "User: Bonjour, je m'appelle Marie et je suis développeuse Python",
        "Assistant: Enchantée Marie ! Bienvenue."
    ]
)

# Consolidate memory (extract facts, summary, knowledge graph, SKOS)
memory, summary = await memorizer.consolidate()
print(f"Summary: {summary}")
print(f"Extracted {len(memory.facts)} facts")

Fact Structure

Each fact contains:

• content: The fact text
• kind: Category (preference, context, skill, event, opinion, demographic)
• timestamp: Extraction date
• duration_validity: Validity period (P1Y, P3M, P7D, etc.)

Knowledge Graph

Concepts stored as RDF triplets:

{
  "concepts": [
    {
      "subject": {"name": "Marie", "type": "person"},
      "predicate": {"name": "has profession", "type": "relation"},
      "object": {"name": "developer", "type": "occupation"}
    }
  ]
}

SKOS Knowledge Base

Ontological concepts following SKOS standard:

{
  "scheme": {"prefLabel": "User Knowledge Database - user_123"},
  "concepts": [
    {
      "prefLabel": "python-developer",
      "definition": "A developer specialized in Python",
      "broader": ["developer"],
      "related": ["programming", "software-engineering"]
    }
  ]
}

Searching

# Search facts
relevant_facts = await memorizer.search_facts("What does the user like?")

# Search SKOS concepts
relevant_concepts = await memorizer.search_skos("development skills")

Storage

Data persisted to ./memories/{user_id}/:

• memory.json - User profile and facts
• knowledge_graph.json - RDF triplets
• skos_knowledge_db.json - SKOS ontology
• {session_id}/conversation.json - Conversation history

Architecture

LLMMemorizer
├── memory/          → User profile + facts (memory.json)
├── knowledge_graph  → RDF triplets (knowledge_graph.json)
├── skos_knowledge_db → SKOS concepts (skos_knowledge_db.json)
└── conversation     → Message history

Dependencies

• Python 3.12+
• pydantic (data models)
• pytest (testing)