Research Digest

Research Digest

Living Document - Auto-updated by Research Agent

Last Updated: 2026-01-24 21:54 UTC

Total Papers Validated: 4

Pending Review: 0

Implementation Tickets: 8 (Created in JIRA)

Table of Contents

1. Summary Dashboard
2. Validated Papers
3. Implementation Backlog
4. Discovery Queue
5. Completed Work
6. Changelog

Summary Dashboard

Research Focus Areas

Validated Papers

V-001: Graph-RAG for Codebases

arXiv: 2601.08773

Key Finding: Deterministic AST-derived graphs outperform LLM-extracted graphs. Multi-hop queries need graph structure.

Integration: Use deterministic parsing over LLM extraction for document knowledge graphs.

V-002: MemRec - Memory-Augmented Recommender

arXiv: 2601.08816

Key Finding: Separation of reasoning from memory management improves flexibility. Async graph updates enable production performance.

Integration: Adopt memory separation pattern for HAM. Use async updates for GraphRAG maintenance.

V-003: GraphFusionSBR - Multi-Channel Graphs

arXiv: 2601.08497

Key Finding: Multi-channel graph fusion (KG + hypergraph + line graph) with MI maximization improves accuracy.

Integration: Extend GraphRAG with multiple channels. Apply MI maximization for channel alignment.

V-004: ACE - Agentic Context Evolution

arXiv: 2601.08747

Key Finding: Metacognitive "retrieve-or-think" decision reduces unnecessary retrievals and token costs.

Integration: Apply orchestrator pattern to research agent and RAG pipeline for efficiency.

Implementation Backlog

Ticket Queue (Created in JIRA)

MD-5131: Memory Separation Pattern

Summary: Implement memory separation pattern - decouple reasoning from memory management

Acceptance Criteria:

• [ ] Memory operations decoupled from query processing
• [ ] Graph updates run asynchronously without blocking
• [ ] <50ms added latency for memory operations
• [ ] 95% test coverage on memory module

MD-5132: Multi-Channel Graph Fusion

Summary: Implement multi-channel graph fusion for enhanced GraphRAG

Acceptance Criteria:

• [ ] Three graph channels operational
• [ ] MI maximization training loop implemented
• [ ] Benchmark: Improvement on multi-hop queries vs single-channel
• [ ] Documentation of channel contributions

MD-5133: Retrieve-or-Think Orchestrator

Summary: Implement metacognitive orchestrator for efficient retrieval decisions

Acceptance Criteria:

• [ ] Orchestrator correctly decides retrieve vs reason 80%+ of time
• [ ] Token usage reduced by 30%+ vs always-retrieve baseline
• [ ] Latency overhead <100ms per decision
• [ ] A/B test framework for orchestrator tuning

MD-5134: Deterministic Entity Extraction

Summary: Replace LLM-based entity extraction with deterministic approach

Acceptance Criteria:

• [ ] Rule-based extraction covers 80%+ of entities
• [ ] Processing time reduced by 10x vs LLM-only
• [ ] Coverage metrics dashboard
• [ ] Fallback to LLM for complex cases

MD-5135: Graph Edge Denoising

Summary: Implement edge pruning and denoising for knowledge graphs

Acceptance Criteria:

• [ ] Edge pruning reduces graph size by 20%+ without accuracy loss
• [ ] Configurable confidence thresholds
• [ ] Automated pruning on graph updates
• [ ] Rollback capability for pruning decisions

MD-5136: Collaborative Signal Integration

Summary: Add collaborative filtering signals to paper recommendations

Acceptance Criteria:

• [ ] Session tracking operational
• [ ] Collaborative scores influence paper ranking
• [ ] Cold-start handling for new users/papers
• [ ] Privacy review completed

MD-5137: Hypergraph Channel

Summary: Implement hypergraph channel for higher-order relations

Acceptance Criteria:

• [ ] Hypergraph operational
• [ ] Higher-order relations captured
• [ ] Integration with fusion layer

MD-5138: Research Agent Token Optimization

Summary: Optimize token usage in research agent operations

Acceptance Criteria:

• [ ] Token budget enforced
• [ ] Cost tracking dashboard
• [ ] 20%+ token reduction achieved

Implementation Roadmap

Phase 1: Foundation ─────────────────────────────────────
│
├─► MD-5133: Retrieve-or-Think Orchestrator [HIGH]
│   └─ Standalone, no dependencies
│
└─► MD-5134: Deterministic Entity Extraction [MEDIUM]
    └─ Enhances existing EntityExtractor

Phase 2: Memory Architecture ────────────────────────────
│
├─► MD-5131: Memory Separation Pattern [HIGH]
│   └─ Requires Phase 1 complete
│
└─► MD-5136: Collaborative Signals [MEDIUM]
    └─ Depends on MD-5131

Phase 3: Graph Enhancement ──────────────────────────────
│
├─► MD-5132: Multi-Channel Graph Fusion [HIGH]
│   └─ Depends on MD-5131
│
├─► MD-5135: Graph Edge Denoising [MEDIUM]
│   └─ Depends on MD-5132
│
└─► MD-5137: Hypergraph Channel [LOW]
    └─ Depends on MD-5132

Phase 4: Optimization ───────────────────────────────────
│
└─► MD-5138: Token Optimization [LOW]
    └─ Depends on MD-5133

Discovery Queue

Papers discovered, pending validation:

New papers will appear here as the daemon discovers them

Completed Work

Changelog

Agent Status

Configuration

Atlassian MCP

{
  "atlassian": {
    "type": "sse",
    "url": "https://mcp.atlassian.com/v1/sse"
  }
}

JIRA Project

• Project Key: MD (Maestro Delivery)
• Cloud ID: 93ba64f7-c16e-4f49-bc8e-43afbc442531
• Issue Type: Task
• Labels: research-agent, implementation

Files

• Validations: research_agent/validations/*.json
• Agent logs: research_agent/daemon.log
• This document: docs/research/RESEARCH_DIGEST.md
• Digest updater: research_agent/digest_updater.py

Living document - Auto-updated by Research Agent
Contact: kulbir.minhas@aiinitiative.co.uk