TL;DR: In October 2025, we successfully integrated Project Continuum (a parallel research intelligence platform design) into Mamut Lab as Layer 11, creating the first universal agentic platform supporting both enterprise orchestration and multi-year research investigations. This post documents the architectural redesign, decision rationale, and what it means for users.

The Strategic Question

For months, we've been developing two parallel platform concepts:

  • Mamut Lab: Enterprise agentic orchestration with Darwin-Godel self-improvement, neurosymbolic reasoning, and coordinated space architecture
  • Project Continuum: Research intelligence platform with temporal knowledge persistence, multi-domain synthesis, and graduated human cognitive partnership

Both architectures shared 90% of their conceptual foundations—neurosymbolic reasoning, event sourcing, dual-process cognition, continual learning. Both targeted knowledge workers who need AI assistance without losing expertise.

The question: Should we maintain two separate products, or integrate them into a unified platform?

The Integration Decision

After extensive architectural analysis (documented in ADR-004), we chose additive integration:

  • Add Layer 11: Research Intelligence to Mamut Lab' 10-layer architecture
  • Enhance existing layers with research-specific patterns (Temporal Knowledge Substrate in Layer 2, multi-model verification in Layer 5, graduated autonomy in Layer 8)
  • Preserve Mamut Lab brand identity, naming conventions, and core architecture
  • Create unified go-to-market: "Universal Agentic Platform" supporting both orchestration and research

Rationale: Single engineering team cannot sustainably maintain two 90%-overlapping platforms. Knowledge workers increasingly need both orchestration (complex workflows) and research (multi-month investigations). Integration provides a unified solution for both use cases.

What Changed: The New 11-Layer Architecture

Layer 11: Research Intelligence (NEW)

The top layer now provides multi-year investigation capabilities:

  • Investigation Lifecycle Management: Initialize, pause/resume, handoff, and conclude research spanning months to years
  • Multi-Domain Synthesis: Unified semantic layer integrating academic papers + market analysis + technical docs + regulatory requirements
  • Investigation Workflows: Specialized maneuvers for literature review, hypothesis testing, gap analysis, and cross-domain impact assessment
  • Workspace Isolation: Rollback-safe exploratory branches with Git-like versioning and automatic cleanup on failures
  • Research-Specific Safety: Hallucination prevention, contradiction management, confidence calibration

Example workflow: PhD researcher investigates quantum computing error correction from January to May, pauses for teaching duties, resumes in September with full context restoration—all active hypotheses, reasoning chains, and pending verifications intact.

Layer 2 Enhancement: Temporal Knowledge Substrate

Memory & Context layer now includes a 4th consolidation tier for long-term research:

  • Versioned Knowledge Graphs: Full temporal provenance (creation metadata, confidence trajectories, contradiction tracking, source lineage)
  • Time-Travel Queries: Access understanding state at any past timestamp ("What did we know about X in March?")
  • Persistent Investigation State: Resumable workflows with active hypotheses, reasoning chains, and verification status
  • Evolutionary Understanding: Meta-knowledge about why decisions were made, what failed and why
  • Dead-End Prevention: Failed exploration paths explicitly tracked to prevent redundant investigation

Implementation: ArangoDB multi-model database (documents + graphs + vectors + time-series) with custom versioning layer extending event sourcing with semantic understanding. Think Git for knowledge graphs.

Layer 5 Enhancement: Multi-Model Verification

Neurosymbolic Reasoning layer gains research-grade verification:

  • Layered Verification Pyramid:
    1. Source Grounding (required for all claims)
    2. Multi-Model Consensus (≥3 models, ≥75% agreement)
    3. Symbolic Validation (ontology consistency checks)
    4. Formal Proof (theorem provers for critical claims)
  • Ensemble Collaboration Protocol: Perception models + LLMs + logic engines + domain-specific models (BioGPT, FinGPT, CodexLLM) improve error detection through ensemble approaches
  • Hallucination Prevention: Source citation requirements, contradiction detection, confidence thresholds, uncertainty quantification

Example: When synthesizing technical claims across academic papers, the system requires ≥3 models to agree on the interpretation before accepting it as fact. Contradictions flagged, uncertainties quantified, sources always traceable.

Layer 8 Enhancement: Graduated Autonomy Framework

Human-AI Collaboration layer replaces "variable autonomy" with explicit 5-level framework preventing automation complacency:

  • Level 0 - Manual Operation: AI disabled for skill maintenance
  • Level 1 - Augmented Assistance: AI suggests, human executes (learning mode)
  • Level 2 - Collaborative Analysis: AI analyzes, human validates reasoning chains
  • Level 3 - Delegated Investigation: AI conducts workflows, human verifies critical decisions
  • Level 4 - Autonomous Operation: AI handles routine, flags surprises for human review

Mandatory Skill Preservation Protocols:

  • Monthly manual mode (Level 0/1) to track skill retention
  • Verification rate monitoring (alert if human verification drops below 40%)
  • Quarterly red team scenarios (error detection exercises)
  • Interleaved practice (vary autonomy levels under cognitive load)

Why this matters: Aviation research shows humans over-trust automation at ~70% reliability (the "trust paradox"). Graduated autonomy with forced skill practice prevents expertise erosion as AI capabilities increase.

What Stayed the Same: Preserving Mamut Lab Identity

The integration is additive, not a replacement:

  • ✅ Mamut Lab brand, naming conventions, visual identity
  • ✅ Coordinated Space Architecture (6 components + 5 spaces)
  • ✅ Darwin-Godel self-improvement (Layer 9)
  • ✅ Cascade prevention (Layer 6)
  • ✅ Continual learning (Layer 7)
  • ✅ Multimodal execution (Layer 4)
  • ✅ Dual-process cognition (Layer 3)
  • ✅ Execution substrate (Layer 1)
  • ✅ Event-driven architecture (CQRS, event sourcing)
  • ✅ Polyglot technology stack (Go/Python/Rust/TypeScript)

For existing users: All orchestration capabilities remain functional. Research Intelligence is an extension—you can use Mamut Lab for enterprise workflows without ever touching Layer 11. But when you need multi-month investigations, the capability is there.

Comparing Approaches: How Mamut Lab Differs

Mamut Lab takes a different architectural approach compared to existing tools:

Compared to AI Research Assistants (Elicit, Perplexity, Consensus)

Feature Research Assistants Mamut Lab Approach
Session Continuity Each session independent Full investigation state persists months/years
Knowledge Evolution Static; papers found don't inform future Cumulative learning; system gets smarter about domain
Reasoning Chains Answer + citations Complete reasoning DAG with provenance
Multi-Domain Siloed (academic OR web) Unified semantic layer (technical + market + scientific + regulatory)
Verification Trust AI or verify manually Neurosymbolic formal verification with symbolic proofs
Skill Preservation No consideration Explicit anti-complacency design with graduated autonomy

Compared to Coding Assistants (GitHub Copilot, Cursor)

Aspect Coding Assistants Mamut Lab Approach
Primary Use Case Code generation Knowledge synthesis → code
Context Window Current file + recent Entire investigation (months/years) with TKS
Verification Manual testing + CI/CD Neurosymbolic formal verification
Domain Breadth Software engineering Technical + Market + Scientific + Regulatory

Platform capabilities: Mamut Lab now combines:

  1. Enterprise orchestration (complex workflows, compliance, audit trails)
  2. Research intelligence (multi-year investigations, temporal knowledge)
  3. Neurosymbolic formal verification (explainable decisions with mathematical guarantees)
  4. Graduated autonomy (skill-preserving human-AI partnership)

Implementation Roadmap: 4 Phases, 12 Months

Phase 1 (Months 1-3): Foundation

  • Temporal Knowledge Substrate data model and storage
  • Basic investigation lifecycle (create, checkpoint, resume)
  • Single-domain literature review maneuver
  • Autonomy Levels 1-2 (Augmented, Collaborative)

Phase 2 (Months 4-6): Multi-Domain

  • Multi-domain synthesis engine
  • Cross-domain reasoning patterns
  • Domain-specific model integration (BioGPT, FinGPT)
  • Contradiction detection and management

Phase 3 (Months 7-9): Advanced Verification

  • Multi-model ensemble verification
  • Symbolic validation layer (Z3, SymPy)
  • Formal proof generation
  • Hallucination prevention protocol

Phase 4 (Months 10-12): Full Autonomy

  • Complete 5-level autonomy framework
  • Skill preservation protocols (monthly manual, red team)
  • Calibrated trust infrastructure
  • Dynamic autonomy adjustment

Target: Q1 2026 for Phase 1, Q3 2026 for limited beta with orchestration + research capabilities.

Documentation: 100+ Files and Counting

The integration is fully documented with transparent architectural rationale:

Total documentation: Expanded from 90+ to 100+ files with Project Continuum integration.

What This Means for Users

For Developers (Original Mamut Lab Users)

  • No breaking changes: All orchestration capabilities remain functional
  • New capability: Long-running technical investigations (feasibility studies, architecture research, library evaluation) now persist across weeks/months
  • Enhanced reasoning: Multi-model verification improves error detection in generated code
  • Skill preservation: Graduated autonomy prevents over-reliance on AI assistance

For Researchers (New User Segment)

  • Multi-year investigations: PhD research, industrial R&D, drug discovery with full context restoration after extended breaks
  • Cross-domain synthesis: Integrate academic papers + market analysis + technical specs + regulatory requirements in unified semantic layer
  • Formal verification: Research claims backed by multi-model consensus and symbolic validation
  • Collaboration: Transfer complete investigation context (reasoning chains, active hypotheses, dead ends) to colleagues

For Technical Leaders

  • Research-first decisions: Investigate technical approaches before committing to architecture
  • Due diligence: VCs and technical strategists can conduct deep-tech startup analysis with temporal knowledge persistence
  • Explainable AI: Neurosymbolic reasoning provides audit trails for critical business decisions
  • Team expertise preservation: Graduated autonomy prevents automation complacency as AI capabilities increase

The Research-First Philosophy

This integration demonstrates the research-first methodology Mamut Lab advocates:

  • We investigated Project Continuum concepts deeply (multi-month research documented in 60+ pages)
  • We evaluated alternatives rigorously (5 integration strategies, documented pros/cons in ADRs)
  • We synthesized cross-domain knowledge (human factors research from aviation + knowledge graph evolution + automation complacency literature)
  • We documented every decision transparently (100+ public docs on GitHub)

Result: Additive integration strategy that preserves brand identity while expanding market positioning—validated through architectural analysis, not marketing slogans.

Looking Ahead: What's Next

Immediate priorities:

  1. Prototype Temporal Knowledge Substrate (ArangoDB implementation with versioning layer)
  2. Build investigation lifecycle manager (pause/resume/handoff)
  3. Implement literature review maneuver (single-domain academic search with quality filtering)
  4. Deploy Levels 1-2 graduated autonomy (Augmented + Collaborative modes)

Long-term vision: Mamut Lab aims to support knowledge work requiring both operational excellence (orchestration) and deep understanding (research)—from software teams investigating technical approaches to PhD researchers conducting multi-year studies to VCs performing technical due diligence.

A platform where you can both coordinate complex workflows and conduct months-long investigations without losing context. Where AI assistance enhances human expertise instead of eroding it. Where every decision is explainable, verifiable, and grounded in formal reasoning.

Get Involved

Track progress: All development happens in the open on GitHub. Watch the repository for updates.

Early access: Interested in beta testing research intelligence workflows? Contact info@Mamut Lab.net with your use case (PhD research, industrial R&D, technical due diligence, etc.).

Feedback: See something that could be improved? Open an issue on GitHub or email suggestions. This is a solo developer project—community input directly shapes priorities.

This integration represents 3 weeks of intensive architectural work: reading Project Continuum design documents, analyzing integration strategies, writing 5 comprehensive ADRs, creating 60+ pages of Layer 11 specification, implementing Temporal Knowledge Substrate pattern documentation, and updating 100+ files. All done transparently with AI assistance (Claude, Copilot)—demonstrating the research-first workflow Mamut Lab aims to provide.

Read the complete integration summary: PROJECT-REDESIGN-SUMMARY.md