Knowledge Forge

Inspiration

Modern AI chatbots are stateless by default — they forget context, repeat questions, and fail to explain why they respond the way they do. We were inspired by this gap between how humans remember (events, preferences, decisions) and how most AI systems operate today. Our goal was to build an AI agent with transparent, persistent memory that can reason over past interactions instead of treating every conversation as a blank slate.

What it does

Our project is an AI agent with structured, persistent memory powered by a graph database.

It:

Stores conversations, user preferences, and actions as a memory graph
Distinguishes between short-term, long-term (episodic), and semantic memory
Uses Neo4j to model relationships between users, messages, topics, and decisions
Allows the agent to explain why it answered something by showing the memory path
Prevents repeated questions and maintains continuity across sessions

How we built it

Neo4j as the core memory store (entities + relationships)
MemMachine-style memory layer to classify and manage memory types
LLM API for reasoning and natural language understanding
Backend (Python) to handle memory writes, reads, and promotion rules
Graph RAG to retrieve structured context instead of flat text chunks
Docker for local deployment and reproducibility

Memory is written after each interaction and selectively retrieved before each response, ensuring relevance without context overload.

Challenges we ran into

Designing a memory schema that avoids duplication and memory bloat
Deciding what should be remembered vs what should be forgotten
Ensuring graph queries stay fast as memory grows
Avoiding hallucinations by grounding responses strictly in stored memory
Managing Docker containers, images, and persistence correctly under time pressure