Context Handover | Preserve Semantic Continuity Across LLM Sessions

The Challenge

The Problem & Solution

LLMs forget everything when a session ends. Standard memory is either too dumb or too expensive.

❌

The Problem

LLMs have no persistent memory across sessions. When a conversation ends, all context is lost. Existing solutions either use linear history (too dumb) or full vector re-indexing (too expensive).

✅

Our Solution

Context Handover extracts Semantic Atoms, measures Context Drift, and optimally packs context using a Bounded Knapsack Algorithm for maximum efficiency.

Capabilities

Core Features

Powerful tools for managing LLM context across sessions

🔍

Semantic Extraction

Extract meaningful context units from conversations using LLM + Regex parsing

📊

Drift Detection

Measure topic shifts in real-time using KL Divergence and Cosine similarity

🎯

Smart Packing

Optimize token usage mathematically with value-density knapsack algorithm

🔗

DAG Tracking

Track dependencies across sessions using directed acyclic graphs

🛡️

Enterprise Ready

Circuit breakers, retries, dead letter queue, and idempotency guarantees

📈

Visual Analytics

Interactive observability dashboard for debugging and monitoring

System Design

Architecture Flow

How Context Handover processes and preserves semantic context

Processing Pipeline

flowchart TD
    subgraph Input["Input Layer"]
        A[User Chat
Raw Text]
    end
    
    subgraph Processing["Processing Layer"]
        B[Atom Extractor
LLM + Regex]
        C[Atom Registry
Dedup + Embed]
        D[Vector Store
Chroma/Qdrant]
    end
    
    subgraph Optimization["Optimization Layer"]
        E[Drift Detector
KL + Cosine]
        F[Token Budgeter
Knapsack Algo]
    end
    
    subgraph Output["Output Layer"]
        G[Loss Ledger
Audit Trail]
        H[New Session
Optimized Prompt]
    end
    
    A --> B
    B --> C
    C --> D
    D --> E
    D --> F
    E --> G
    F --> H
    
    style Input fill:#1a1a1a,stroke:#c41e3a,color:#fff
    style Processing fill:#1a1a1a,stroke:#8b1528,color:#fff
    style Optimization fill:#1a1a1a,stroke:#c41e3a,color:#fff
    style Output fill:#1a1a1a,stroke:#8b1528,color:#fff
    style A fill:#2a1a1a,stroke:#c41e3a,color:#fff
    style B fill:#2a1a1a,stroke:#c41e3a,color:#fff
    style C fill:#2a1a1a,stroke:#c41e3a,color:#fff
    style D fill:#2a1a1a,stroke:#c41e3a,color:#fff
    style E fill:#2a1a1a,stroke:#c41e3a,color:#fff
    style F fill:#2a1a1a,stroke:#c41e3a,color:#fff

Data Flow

graph LR
    A[User Chat] --> B[Atom Extractor]
    B --> C[Atom Registry]
    C --> D[Vector Store]
    D --> E[Drift Detector]
    D --> F[Token Budgeter]
    E --> G[Loss Ledger]
    F --> H[New Session]
    
    style A fill:#1a1a1a,stroke:#ff2d4f,color:#fff
    style B fill:#1a1a1a,stroke:#c41e3a,color:#fff
    style C fill:#1a1a1a,stroke:#8b1528,color:#fff
    style D fill:#1a1a1a,stroke:#c41e3a,color:#fff
    style E fill:#1a1a1a,stroke:#ff2d4f,color:#fff
    style F fill:#1a1a1a,stroke:#8b1528,color:#fff
    style G fill:#1a1a1a,stroke:#c41e3a,color:#fff
    style H fill:#1a1a1a,stroke:#ff2d4f,color:#fff

Fundamentals

Key Concepts

🔷

Semantic Atom

The smallest unit of meaningful context. Like a single Lego brick that can be recombined.

🌳

Session DAG

Directed Acyclic Graph tracking atom relationships across sessions like a family tree.

🧭

Drift Metric

Measures topic change since last handover, like a compass checking your course.

🎒

Knapsack Budget

Mathematically selects the most valuable atoms for the token limit, like packing a suitcase.

Getting Started

How It Works

Three simple steps to preserve context across sessions

Initialize & Add Messages

Create a SessionManager and add your conversation messages. The system automatically tracks context as it's built.

from context_handover import SessionManager

manager = SessionManager(session_id="session_001")
manager.add_message(role="user", content="I want to build a rocket engine using methane.")
manager.add_message(role="assistant", content="Understood. Methane (CH4) offers high specific impulse...")

Extract Semantic Atoms

Automatically extract meaningful context units from your conversation history. Atoms capture semantic meaning, not just raw text.

# Extract atoms automatically
atoms = manager.extract_atoms()
print(f"Extracted {len(atoms)} semantic atoms.")

Build Handover Package

Create an optimized context package for the new session. The Knapsack algorithm ensures maximum value within token limits.

# Handover to a new session (preserving context)
new_session_pkg = manager.build_handover_package()
manager.handover_to_new_session("session_002", new_session_pkg)

Approach	Token Efficiency	Semantic Coherence	Latency	Auditability
❌ Naive Buffer	Low	Low	Fast ✅	None ❌
⚡ Vector Recall	Medium	Medium	Slow ~200ms ⚠️	Low ⚠️
✅ Context Handover	High 🎯	High 🎯	Fast ~40ms ⚡	Full ✅

Ecosystem

Works With Your Stack

Seamless integration with popular frameworks

Integration Architecture

flowchart LR
    A[Context Handover] --> B[LangChain]
    A --> C[LlamaIndex]
    A --> D[AutoGen]
    A --> E[LangGraph]
    A --> F[OpenTelemetry]
    A --> G[Langfuse]
    
    style A fill:#c41e3a,color:white
    style B fill:#1a1a1a,stroke:#2196F3,color:#fff
    style C fill:#1a1a1a,stroke:#FF9800,color:#fff
    style D fill:#1a1a1a,stroke:#9C27B0,color:#fff
    style E fill:#1a1a1a,stroke:#FF5722,color:#fff
    style F fill:#1a1a1a,stroke:#00BCD4,color:#fff
    style G fill:#1a1a1a,stroke:#E91E63,color:#fff

⛓️

LangChain

Custom Memory Module

📄

LlamaIndex

Node Parser

🤖

AutoGen

State Handover

📊

OpenTelemetry

Native Tracing

📈

Langfuse

Observability

Enterprise

Production Ready

Built for reliability at scale

🔄

Idempotency

Duplicate events auto-detected & ignored

🔁

Smart Retries

Exponential backoff for LLM/Redis failures

⚡

Circuit Breakers

Prevents cascading failures

📦

Dead Letter Queue

Failed events saved for replay

🔒

PII Ready

Redaction & encryption hooks

📊

Full Auditability

Complete loss ledger for compliance

Workflow

Semantic Atom Lifecycle

How atoms move through the system

sequenceDiagram
    participant User
    participant Manager
    participant Extractor
    participant Registry
    participant Vector
    
    User->>Manager: Add Message
    Manager->>Extractor: Trigger Extraction
    Extractor->>Extractor: Parse LLM + Regex
    Extractor->>Registry: Submit Atoms
    Registry->>Registry: Deduplicate
    Registry->>Vector: Generate Embeddings
    Vector-->>Manager: Confirmation
    Manager-->>User: Done
    
    style User fill:#1a1a1a,stroke:#c41e3a,color:#fff
    style Manager fill:#1a1a1a,stroke:#ff2d4f,color:#fff
    style Extractor fill:#1a1a1a,stroke:#8b1528,color:#fff
    style Registry fill:#1a1a1a,stroke:#c41e3a,color:#fff
    style Vector fill:#1a1a1a,stroke:#ff2d4f,color:#fff

Example

Quick Start

Get up and running in minutes

                    context_handover_example.py
                    Python
                

                    from context_handover import SessionManager, SemanticAtom

# Initialize manager
manager = SessionManager(session_id="session_001")

# Add meaningful interactions
manager.add_message(
    role="user", 
    content="I want to build a rocket engine using methane."
)
manager.add_message(
    role="assistant", 
    content="Understood. Methane (CH4) offers high specific impulse..."
)

# Extract atoms automatically
atoms = manager.extract_atoms() 
print(f"Extracted {len(atoms)} semantic atoms.")

# Handover to a new session
new_session_pkg = manager.build_handover_package()
manager.handover_to_new_session("session_002", new_session_pkg)
                

Preserve Semantic ContinuityAcross LLM Session Boundaries