sinapse-ai 9.3.0 → 9.4.0

package/squads/squad-research/knowledge-base/agentic-second-brain-reference.md

@@ -0,0 +1,591 @@
+# Agentic Second Brain Engineering Reference
+
+Comprehensive reference on building agentic Second Brains -- systems where AI agents capture, organize, connect, and retrieve knowledge autonomously, transforming personal vaults into living cognitive infrastructure.
+
+**Sources:** 67 sources (papers, repositories, articles, documentation). Research conducted April 2026.
+
+---
+
+## 1. The Convergence
+
+Three waves converge to create the Agentic Second Brain:
+
+1. **PKM Wave:** Obsidian, Logseq, Roam Research -- local-first vaults with markdown and bidirectional links
+2. **LLM Wave:** GPT-4, Claude, Gemini -- models capable of reading, summarizing, synthesizing, and producing knowledge
+3. **Agentic Wave:** CrewAI, LangGraph, Claude Code, OpenAI Agents SDK -- frameworks giving agents autonomy, tools, and persistent memory
+
+### Why Now (2025-2026)
+
+- Context windows of 1M+ tokens (Claude, Gemini) allow agents to read entire vaults
+- MCP (Model Context Protocol) enables LLM-to-filesystem connections
+- A-Mem (NeurIPS 2025) formalized agentic memory inspired by Zettelkasten
+- Graphiti/Zep: temporal knowledge graphs tracking how facts change
+- Context Engineering recognized formally (Karpathy, Willison, 2025) as more important than prompt crafting
+
+### Market Context
+
+- Graph database market: $2.85B (2025), projected $18-20B by 2032-2034
+- Knowledge management market: $773B-$885B (2024)
+- Gartner: 33% of enterprise apps will include agentic AI by 2028 (vs <1% in 2024)
+- Cost of poor data quality: $12.9M/year per organization
+
+---
+
+## 2. Knowledge Architecture (System 1)
+
+### What and Why
+
+Knowledge Architecture is the formal design that organizes, classifies, and relates units of knowledge. Without it, a vault becomes a graveyard of notes -- information accumulated but inaccessible. Architecture determines findability, connectability, evolution, and reasoning capability.
+
+### Frameworks
+
+| Framework | Structure | Focus | Ideal For |
+|-----------|-----------|-------|-----------|
+| Zettelkasten | Network of atomic interlinked notes | Idea emergence | Research, writing |
+| PARA | Projects / Areas / Resources / Archives | Action and productivity | Project management |
+| Evergreen Notes | Conceptual notes that evolve | Durable thinking | Deep reflection |
+| MOC (Maps of Content) | Index notes aggregating themes | Navigation | Large vaults |
+| Knowledge Graph | Entities + relations + attributes | Reasoning and inference | Agentic systems |
+
+### State of the Art
+
+**GraphRAG:** Evolution of traditional RAG incorporating a knowledge graph in retrieval. Instead of pure vector search, the system also queries the graph for relevant entities and relations, combining structural precision with semantic coverage.
+
+**Graphiti (Zep):** Framework for temporal context graphs. Unlike static graphs, Graphiti tracks how facts change over time, maintains provenance for data sources, and supports both prescribed and learned ontology.
+
+**Bi-Temporal Model:** Tracks when an event occurred AND when it was ingested. Each edge includes explicit validity intervals. Conflicts use temporal metadata to update or invalidate -- but not discard -- outdated information.
+
+### Implementation Steps
+
+1. Define base ontology: entity types (Person, Concept, Project, Decision, Insight, Source) and relations
+2. Choose granularity: atomic notes (1 idea = 1 note) vs composite notes
+3. Implement layers: fixed taxonomy + fluid tags + bidirectional links
+4. Automatic graphs: use LLMs to extract entities and relations
+5. Evolutionary ontology: allow the system to learn new categories from usage
+
+### Risks
+
+- Over-engineering ontologies nobody can maintain
+- Rigid taxonomies that reject emergent knowledge
+- Platform lock-in from tool-specific architectures
+- Temporal inconsistency polluting the graph without invalidation mechanisms
+
+### Key People
+
+- **Niklas Luhmann** -- Zettelkasten creator (70 books published)
+- **Tim Berners-Lee** -- Semantic Web, RDF, OWL foundations
+- **Preston Rasmussen** -- Zep/Graphiti temporal graphs paper
+
+---
+
+## 3. Ingestion, Capture and Context Normalization (System 2)
+
+### The Problem
+
+Most valuable knowledge is generated in ephemeral contexts -- a conversation with Claude, a meeting, an article read quickly. Without systematic capture, knowledge evaporates.
+
+### State of the Art
+
+**Agentic Context Engineering (ACE):** Academic framework with three specialized components -- Generator, Reflector, and Curator -- representing context as structured bullets instead of monolithic prompts. Each memory entry has metadata (unique IDs, utility counters) and content capturing reusable strategies, domain concepts, or common failure modes.
+
+**Cognee:** Open-source cognitive memory engine combining graph structures with vector embeddings. Modular pipelines for custom extraction, enrichment, and retrieval. 30+ connectors for documents, images, audio, conversations. $7.5M seed (Feb 2026), 1M+ pipeline runs, 70+ companies in production.
+
+### Source Types
+
+| Source | Content Type | Normalization Challenge |
+|--------|-------------|------------------------|
+| Claude/ChatGPT conversations | Dialogs, decisions, insights | Extract essence from conversational flow |
+| Documents (PDF, Word, Slides) | Structured text + tables | OCR, layout parsing, table extraction |
+| Web pages | Articles, docs, forums | HTML cleanup, main content extraction |
+| Audio/Video | Meetings, podcasts | Transcription + diarization + summary |
+| Source code | Logic, patterns, architecture | Intent extraction beyond syntax |
+| Manual notes | Drafts, brainstorms | Structuring and connecting |
+
+### Normalization Pipeline
+
+```
+[Raw Source]
+  -> Extraction (parsing, OCR, transcription)
+  -> Cleaning (remove noise, format)
+  -> Semantic Chunking (divide by meaning, not tokens)
+  -> Enrichment (metadata, tags, entities)
+  -> Embedding (vectorization for semantic search)
+  -> Indexing (vault + graph + vector store)
+  -> Connection (links with existing knowledge)
+```
+
+### Implementation
+
+1. One connector per source (API, filesystem watch, clipboard monitor, webhook)
+2. Semantic chunking: divide by meaning, each chunk self-contained
+3. Entity extractors: LLMs identify people, concepts, decisions, actions
+4. Deduplication: verify knowledge exists before creating new note
+5. Templates per type: meeting note, article digest, conversation insight
+6. Auto-connection: agent identifies related existing notes and creates links
+
+### Risks
+
+- Ingestion without curation generates noise degrading vault quality
+- Loss of provenance makes knowledge unverifiable
+- Over-chunking destroys context necessary for comprehension
+
+---
+
+## 4. Operational Memory and Context Engineering (System 3)
+
+### The Paradigm Shift
+
+Andrej Karpathy (2025): "Context engineering is the delicate art and science of filling the context window with just the right information for the next step."
+
+Simon Willison: "Context engineering is what we do instead of fine-tuning."
+
+Mental model: think of the LLM as a CPU and the context window as RAM. Your job is analogous to an operating system: load working memory with exactly the right code and data for the task.
+
+### Memory Layers
+
+| Layer | Human Analogy | Function | Duration | Cost |
+|-------|--------------|----------|----------|------|
+| Working Memory | Scratchpad | Active task information | Session | Direct tokens |
+| Episodic Memory | Autobiographical | Specific events, past interactions | Days-Months | Vector search |
+| Semantic Memory | General knowledge | Facts, concepts, rules | Permanent | Graph + Vector |
+
+### Storage Tiers
+
+| Tier | Access | Example |
+|------|--------|---------|
+| HOT | In active context window | System prompt, current task, recent messages |
+| WARM | Retrievable in <300ms | Recent notes, mentioned entities, project decisions |
+| COLD | Retrievable on demand | Old notes, completed projects, full history |
+
+**ContextForge principle:** Memory becomes HOT not just when accessed, but when semantically crucial for the current task. Relevance pulls it to HOT instantly regardless of age.
+
+### Token Budget Management
+
+```
+TOTAL WINDOW: 200,000 tokens
+  |- System Prompt:      ~2,000 (1%)
+  |- Agent Persona:      ~1,500 (0.75%)
+  |- Memory Context:    ~50,000 (25%)     <- MANAGED
+  |- User History:      ~20,000 (10%)     <- COMPACTED
+  |- Tool Results:      ~50,000 (25%)     <- DYNAMIC
+  |- Response Budget:   ~76,500 (38.25%)  <- RESERVED
+```
+
+### Frameworks and Tools
+
+| Tool | Approach | Highlight |
+|------|----------|-----------|
+| **Letta (MemGPT)** | Virtual context management (OS-inspired) | #1 Terminal-Bench. Conversations API for cross-session shared memory |
+| **Mem0** | Memory layer with graph DB | 26% accuracy gain vs OpenAI Memory, 91% less latency, 90% less tokens. $24M Series A |
+| **A-Mem** | Zettelkasten-inspired agentic memory | NeurIPS 2025. Atomic notes with keywords, tags, contextual descriptions |
+| **OpenMemory** | Local-first persistent memory | SQL-native with temporal graphs and entity tracking |
+| **ContextForge** | Three-tier memory system | Auto promotion/demotion. 10x compression ratio |
+| **Hindsight** | Four separate memory networks | 91.4% accuracy on DMR benchmark. Open-source |
+| **Cognee** | Graph-vector hybrid cognitive engine | $7.5M seed, 1M+ pipeline runs, 30+ data connectors |
+
+### Risks
+
+- Context pollution degrades response quality
+- Lost-in-the-middle effect with long contexts
+- Memory staleness presenting outdated info as current
+- Over-compaction losing critical nuances
+
+---
+
+## 5. Retrieval, Navigation and Discovery (System 4)
+
+### Hybrid Search as Production Standard (2026)
+
+```
+[User Query]
+  -> BM25 (keyword search) -> Top-K results
+  -> Dense Embeddings (semantic) -> Top-K results
+  -> Knowledge Graph (structured) -> Entities/Relations
+  -> Reciprocal Rank Fusion (RRF) -> Merged and Ranked
+  -> Cross-Encoder Reranking -> Final Top-N
+  -> LLM Generation with Context
+```
+
+**Why BM25 still matters:** Unbeatable for product codes, legal terminology, unique acronyms. For most real RAG applications, full-text search provides precision that vector search alone cannot.
+
+### Vector Database Comparison (2026)
+
+| Database | Ideal For | Scale | Latency (P95) |
+|----------|----------|-------|----------------|
+| Pinecone | Production enterprise | Billions | <50ms |
+| Weaviate | Hybrid search native | Hundreds of millions | <100ms |
+| Qdrant | Performance/cost | Hundreds of millions | <100ms |
+| Chroma | Rapid prototyping | Millions | Variable |
+| pgvector | PostgreSQL integration | 5-100M | Variable |
+| Milvus | Cost at scale | Billions | <50ms |
+
+Strategy: Start with pgvector/Chroma for prototype, migrate to Pinecone/Weaviate for production.
+
+### GraphRAG
+
+Experiments report gains in faithfulness, response relevance, and context recall when integrating structured graph contexts AND dense embeddings:
+
+- **Graph:** High-precision evidence, structured by domain
+- **Vector:** Nuanced context coverage and edge cases
+- **Combined:** Governance via graph + flexibility via vectors
+
+### Zep/Graphiti: Temporal Retrieval
+
+Retrieval P95 of 300ms through hybrid search combining semantic embeddings, keyword (BM25), and direct graph traversal -- no LLM calls during retrieval. 94.8% on DMR benchmark with gpt-4-turbo. Up to 18.5% accuracy gain with 90% latency reduction.
+
+**LazyGraphRAG (Microsoft):** Indexing at vector RAG cost (0.1% of full GraphRAG) with comparable quality for global queries.
+
+### Implementation
+
+1. Dual index: BM25 (full-text) + vector (embeddings) per note
+2. Graph layer: entity/relation graph for structured navigation
+3. Hybrid fusion: RRF or learned merging
+4. Reranking: cross-encoder as second phase
+5. Context window packing: select and order chunks for maximum utility
+6. Feedback loop: track which results are actually used
+
+---
+
+## 6. Agent, Subagent and Skills Modeling (System 5)
+
+### Reasoning Patterns
+
+| Pattern | Description | Ideal Use |
+|---------|-------------|-----------|
+| ReAct | Reason + Act in loop | Tasks with tools |
+| Tree of Thought | Explore multiple reasoning paths | Problems with multiple solutions |
+| Graph of Thought | Reasoning as graph, merge/refine thoughts | Complex multi-source synthesis |
+| Chain of Thought | Step-by-step linear reasoning | Sequential problems |
+| Reflection | Agent evaluates own output | Quality and self-correction |
+
+### Multi-Agent Orchestration Frameworks (2026)
+
+| Framework | Architecture | Control | Maturity |
+|-----------|-------------|---------|----------|
+| LangGraph | State machine with directed graph | Maximum | High |
+| CrewAI | Role-playing + task delegation | Medium | High |
+| OpenAI Agents SDK | Agents with handoffs and guardrails | Medium | High |
+| Claude Agent SDK | Claude-native with tool use | Medium | High |
+| Google ADK | Agent Development Kit | Medium | High |
+| Microsoft Agent Framework | Unifies AutoGen + Semantic Kernel | Medium-High | New (2026) |
+| deepagents (LangChain) | Batteries-included agent harness | Medium | High |
+
+**Note:** Microsoft retired AutoGen in favor of the new Microsoft Agent Framework. AutoGen remains in maintenance mode; the community fork AG2 continues independently.
+
+### Architecture for Second Brain
+
+```
+[Main Orchestrator]
+  |-- [Capture Agent]     -- Monitor sources, ingest
+  |-- [Curation Agent]    -- Connect, tag, classify
+  |-- [Research Agent]    -- Search, navigate, discover
+  |-- [Synthesis Agent]   -- Summarize, combine, produce
+  |-- [Quality Agent]     -- Validate, score, suggest
+  |-- [Maintenance Agent] -- Detect decay, archive, clean
+```
+
+### Skills as Capability Units
+
+Skills are modular capabilities agents can invoke:
+- `/capture` -- Capture insight from conversation
+- `/connect` -- Find and create links between related notes
+- `/remember` -- Retrieve relevant knowledge from vault
+- `/synthesize` -- Combine multiple notes into synthesis
+- `/pipeline` -- Execute complete research pipeline
+- `/graph` -- Visualize knowledge connections
+
+### Implementation
+
+1. Agent-per-concern, not one super-agent
+2. Shared state via state graph or shared memory
+3. Tool registry: catalog of available tools per agent
+4. Handoff protocol: formal transfer between agents
+5. Skill system: atomic functions agents can compose
+6. Guardrails: authority limits per agent
+
+---
+
+## 7. Automations, Hooks and Operational Pipelines (System 6)
+
+### Automation Types
+
+| Type | Trigger | Example |
+|------|---------|---------|
+| Event Hook | Specific action occurs | End of Claude conversation -> capture insights |
+| Cron Job | Temporal schedule | 11pm BRT daily -> review day's sessions |
+| Watch | Filesystem change | New file in folder -> automatic ingestion |
+| Webhook | External HTTP request | GitHub push -> update project notes |
+| Threshold | Metric exceeds limit | Note not accessed for 90 days -> flag for review |
+| Pipeline | Transformation chain | Conversation -> chunks -> entities -> notes -> links |
+
+### Conversation-to-Knowledge Pipeline
+
+```
+[Claude Code session ends]
+  -> Hook: PreCompact captures session digest
+  -> Cron (11pm BRT): Haiku reviews all day's sessions
+  -> Extract: decisions, insights, facts, actions
+  -> For each insight:
+      -> Check if already exists in vault
+      -> If new: create note with appropriate template
+      -> If existing: enrich note with new context
+      -> Create links with related notes
+  -> Update indices and graphs
+  -> Generate daily summary
+```
+
+### Automation Patterns for Second Brain
+
+| Pattern | Frequency | Agent | Action |
+|---------|-----------|-------|--------|
+| Daily Digest | 1x/day | Capture | Review sessions, extract insights |
+| Connection Discovery | Continuous | Curation | Find links between new and existing notes |
+| Decay Detection | 1x/week | Maintenance | Identify outdated notes |
+| Graph Update | Continuous | Curation | Keep knowledge graph synchronized |
+| Quality Scoring | 1x/week | Quality | Score notes by completeness, connections, relevance |
+| Monthly Review | 1x/month | Synthesis | Generate knowledge retrospective |
+
+---
+
+## 8. Governance, Quality and Vault Evolution (System 7)
+
+### Content Decay Types
+
+| Type | Example | Detection |
+|------|---------|-----------|
+| Factual | Framework version changed | Compare with external sources |
+| Contextual | Project was cancelled | Status tracking |
+| Relevance | Topic no longer priority | Usage analytics |
+| Structural | Broken links, obsolete tags | Graph validation |
+| Temporal | Note has date but no review deadline | Metadata check |
+
+### Lifecycle Management
+
+```
+[DRAFT] -> [REVIEWED] -> [PUBLISHED] -> [EVERGREEN]
+                              |               |
+                              v               v
+                        [NEEDS UPDATE]  [DEPRECATED]
+                              |               |
+                              v               v
+                        [UPDATED]       [ARCHIVED]
+```
+
+### Quality Scoring
+
+| Dimension | Weight | Metrics |
+|-----------|--------|---------|
+| Completeness | 20% | All sections filled? Links present? |
+| Accuracy | 25% | Verifiable facts? Sources cited? |
+| Connection | 15% | Number and quality of links |
+| Freshness | 20% | Time since last review, access frequency |
+| Utility | 20% | Retrieval frequency, user feedback |
+
+**Score interpretation:**
+- >= 80: Evergreen (high confidence)
+- 60-79: Healthy (periodic review)
+- 40-59: Needs Attention (flag for review)
+- < 40: At Risk (archival candidate)
+
+### Implementation
+
+1. Mandatory metadata: created_date, updated_date, review_date, status, owner, quality_score
+2. Review cycles by content type (technical notes: quarterly, decisions: semi-annual, facts: continuous)
+3. Automated scoring by quality agent
+4. Decay alerts when notes fall below threshold
+5. Git for entire vault -- every change traceable
+6. Provenance chain: each note knows its origin
+
+---
+
+## 9. Knowledge Analytics and Cognitive Observability (System 8)
+
+### Telemetry Metrics
+
+| Metric | What It Measures | Why It Matters |
+|--------|-----------------|----------------|
+| Retrieval Hit Rate | % of searches returning useful results | Index quality |
+| Knowledge Coverage | % of topics covered vs demanded | Vault gaps |
+| Note Utilization | Access frequency per note | Content relevance |
+| Connection Density | Links per note (avg and distribution) | Network richness |
+| Decay Rate | % of notes becoming outdated per period | Vault health |
+| Ingestion Velocity | Notes created per day/week | Capture pace |
+| Synthesis Rate | Synthetic notes vs raw notes | Knowledge maturity |
+| Agent Token Usage | Tokens consumed per operation | Operational cost |
+| Query Latency | Retrieval response time | Performance |
+| Context Relevance | Relevance score of retrieved context | RAG quality |
+
+### Observability Tools (2025-2026)
+
+| Tool | Focus |
+|------|-------|
+| LangWatch | LLM observability -- traces, evaluations, guardrails |
+| Datadog LLM Obs | Enterprise integration with existing stack |
+| Langfuse | Open-source -- traces, scores, datasets |
+| Arize Phoenix | ML observability -- drift detection, embeddings |
+| Monte Carlo | Data observability -- quality for RAG |
+
+---
+
+## 10. Research, Synthesis and Knowledge Production (System 10)
+
+### Deep Research (2025-2026)
+
+- **OpenAI Deep Research:** Retrieves, reads, critiques, synthesizes hundreds of papers in under an hour
+- **Google Gemini Deep Research:** Synthesizes from dozens of independent sources
+- **The AI Scientist (Sakana AI):** Creates research ideas, writes code, runs experiments, writes manuscripts, performs peer review. Published in Nature (March 2026)
+
+### Research Pipeline
+
+```
+[Question/Topic]
+  -> Decompose into sub-questions
+  -> Parallel search across multiple sources
+  -> For each source:
+      -> Retrieval (web, vault, papers)
+      -> Claim/fact extraction
+      -> Credibility evaluation
+  -> Cross-reference between sources
+  -> Identify convergences and divergences
+  -> Structured synthesis
+  -> Citation verification
+  -> Formatted output (note, article, report)
+```
+
+### Synthesis Methods
+
+| Method | Description | When to Use |
+|--------|-------------|-------------|
+| Aggregation | Combine facts from multiple sources | Topic overview |
+| Comparison | Contrast approaches/perspectives | Decisions between alternatives |
+| Narrative | Build coherent story from facts | Communication and teaching |
+| Framework | Extract mental model from data | Structural comprehension |
+| Gap Analysis | Identify what is MISSING | Direct future research |
+| Contradiction | Identify conflicts between sources | Validation and verification |
+
+### Research Tools (2026)
+
+| Tool | Type | Highlight |
+|------|------|-----------|
+| Elicit | Paper discovery + synthesis | 138M papers, 99.4% screening accuracy |
+| Consensus | Evidence-based answers | Scientific corpus search |
+| Scite.ai | Citation analysis | Verify if papers were supported/contrasted |
+| Perplexity | General research | Web search + inline citations |
+
+---
+
+## 11. Infrastructure, Portability and Resilience (System 11)
+
+### Core Principles
+
+| Principle | Implementation |
+|-----------|---------------|
+| Local-First | Filesystem local as source of truth |
+| Plain Text | Markdown as universal format |
+| Git-Versioned | Full change history traceable |
+| Cloud-Synced | Backup as secondary layer |
+| Portable | No specific runtime dependencies |
+| Resilient | 3-2-1 backup strategy |
+
+### Filesystem Design
+
+```
+vault/
+  00-inbox/              # Quick capture, unprocessed
+  01-daily/              # Daily notes, session logs
+  02-projects/           # Active project notes
+  03-areas/              # Ongoing responsibility areas
+  04-resources/          # Reference material
+  05-archive/            # Archived projects and notes
+  06-templates/          # Templates for new note types
+  07-agents/             # Agent configuration and memory
+  08-analytics/          # Metrics and vault health reports
+  _attachments/          # Images, PDFs, media
+  _data/                 # Structured data (JSON, YAML)
+```
+
+### Backup Strategy (3-2-1)
+
+| Layer | Medium | Frequency | Retention |
+|-------|--------|-----------|-----------|
+| L1 | Git local (commits) | Every change | Unlimited |
+| L2 | Cloud sync (OneDrive) | Real-time | 30 days of versions |
+| L3 | Git remote (GitHub/private) | Daily push | Unlimited |
+| L4 | External SSD (encrypted) | Weekly | Unlimited |
+
+---
+
+## 12. Orchestration for Obsidian + Claude Code (System 12)
+
+### Why This Stack
+
+Obsidian is the most mature and extensible local-first vault tool. Claude Code is the AI agent with the greatest capability to operate on local filesystems. The combination is, in 2026, the most powerful stack for Agentic Second Brains.
+
+### Key Integrations (2025-2026)
+
+**Claudian:** Obsidian plugin embedding AI agents (Claude Code, Codex) as vault collaborators. Vault becomes the agent's working directory with full agentic capabilities. ~4K GitHub stars.
+
+**Obsidian Agent Skills (Jan 2026):** Official repository by Obsidian CEO (Steph Ango/Kepano) teaching AI agents to correctly handle wikilinks, frontmatter, Bases databases, and JSON Canvas. Obsidian 1.12 introduced a native CLI dramatically facilitating agent integration.
+
+**obsidian-claude-pkm (Starter Kit):** Complete kit with 4 specialized agents with memory, 10 skills, auto-commit hooks, and agents that use memory to learn user patterns between sessions.
+
+### CLAUDE.md as Vault Constitution
+
+Claude Code reads CLAUDE.md at project root every session -- analogous to the SINAPSE Constitution. For Second Brain vaults, this file defines:
+
+- Vault structure and conventions
+- Agent permissions and boundaries
+- Capture and curation rules
+- Quality standards
+- Automation triggers
+
+### Human-Machine Interface
+
+| Interface | Function | Complexity |
+|-----------|----------|-----------|
+| Quick Capture | Rapid idea/insight input | Minimal |
+| Search Bar | Semantic + keyword search | Low |
+| Note Editor | Edit with connection preview | Medium |
+| Graph View | Knowledge network visualization | Medium-High |
+| Dashboard | Health metrics, gaps, trends | Medium |
+| Agent Chat | Natural conversation with vault | Low |
+| Command Palette | Skills and automations by command | Low |
+
+### Progressive Disclosure for Second Brain UX
+
+1. **Level 1:** Quick capture (one text box, one button)
+2. **Level 2:** Note created with basic metadata and suggested connections
+3. **Level 3:** Full editing, tags, manual links, template selection
+4. **Level 4:** Graph view, analytics, bulk operations
+
+---
+
+## 13. Key People and References
+
+### Researchers and Practitioners
+
+| Person | Contribution |
+|--------|-------------|
+| Niklas Luhmann | Zettelkasten creator, proved knowledge architecture scales |
+| Tiago Forte | Popularized "Building a Second Brain" (BASB) with PARA method |
+| Andy Matuschak | Evergreen Notes concept |
+| Andrej Karpathy | Defined Context Engineering (2025) |
+| Simon Willison | "Context engineering is what we do instead of fine-tuning" |
+| Vannevar Bush | Memex concept (1945) |
+| Doug Engelbart | "Augmenting Human Intellect" framework (1962) |
+
+### Key Tools and Frameworks
+
+| Tool | Category | Differentiator |
+|------|----------|---------------|
+| Obsidian | PKM vault | Plugin ecosystem, graph view, markdown, local-first |
+| Graphiti/Zep | Temporal graph | Bi-temporal model, provenance tracking |
+| Mem0 | Memory layer | 26% accuracy gain, 90% token reduction |
+| Letta (MemGPT) | Virtual context | OS-inspired, #1 Terminal-Bench |
+| Cognee | Cognitive engine | Graph-vector hybrid, 30+ connectors |
+| ContextForge | Three-tier memory | Auto promotion/demotion, 10x compression |
+
+---
+
+*Knowledge base compiled from research conducted by Scope (Research Analyst) -- SINAPSE Research Initiative*
+*Research: MS-009 Agentic Second Brain Engineering | 67 sources | April 2026*