AI Research Agent with Advanced Memory System

Problem

In long-running research tasks, AI agents suffer from context window overflow, forgetting crucial past findings, and difficulty in resolving contradicting research claims over time.

Architecture

This project implements a multi-layered memory system using a 5-module agentic workflow:

1. Extraction: Raw conversation and tools outputs distilled into candidate facts.
2. Operations: Conflict resolution, contradiction detection, confidence scoring, and memory decay.
3. Persistent Memory: Combined Vector (Qdrant), Keyword (BM25), and Graph (NetworkX) indices.
4. Retrieval & Fusion: Cognitive retrieval combining multi-index searches and ranking.
5. Generation: Citation-backed answer formulation.

Novelty

• Dual-rate decay formula distinguishing episodic (rapid decay) from semantic (consolidated facts) memories.
• Structural contradiction detector comparing research claims using confidence scores instead of simple timestamps.

Benchmarks

Evaluated against LoCoMo/LongMemEval subsets. See benchmarks/results.json for detailed metrics.

Roadmap

• Implement contradiction detector comparing research assertions.
• Connect custom tools and MCP servers for live research retrieval.
• Optimize fusion scoring weight grid search.