Every URL researched as inspiration for PKA's knowledge architecture
Date: 2026-04-15 | MC-602
Most influential single reference. PKA independently built the same 3-layer architecture (raw sources → wiki layer → schema layer). Formally adopted in MC-605.
| URL | Description | Finding / Impact |
|---|---|---|
| gist.github.com/karpathy/442a6bf... | Karpathy LLM Wiki Gist | Personal knowledge system: raw sources → wiki → schema. PKA had 12/12 core patterns already implemented. adopted |
| x.com/karpathy/status/203980... | Karpathy Tweet Thread | Original announcement of the LLM Wiki concept. Validated PKA's approach. adopted |
Evaluated March 2026. Five repos reviewed for patterns and skills to adopt.
| URL | Description | Finding / Impact |
|---|---|---|
| github.com/bradautomates/second-brain | Brad's Second Brain Template | Brain dump workflow, auto-commit hooks, delegation patterns. Adopted: /brain skill, "grep before glob" principle, auto-commit, git-as-database. adopted |
| github.com/earlyaidopters/second-brain | Early Adopters Second Brain | Setup scripts, Gemini file import, beginner template. Less relevant for PKA's maturity level. reviewed |
| github.com/kepano/obsidian-skills | Kepano's Obsidian Skills (Obsidian CEO) | 5 skills: obsidian-markdown, obsidian-bases, json-canvas, obsidian-cli, defuddle. All installed globally. adopted |
| github.com/coleam00/second-brain-skills | Cole's Second Brain Skills | MCP client wrapper, PPTX generation, brand voice. Influenced "progressive disclosure" pattern (metadata in context, body on trigger). influenced |
| github.com/Abilityai/cornelius | Zettelkasten Automation | FAISS search, brain swapping. Too complex; A-MEM paper covers concepts better. reviewed |
Evaluated April 2026 (MC-539). Eight production-grade memory systems compared against PKA's approach.
| URL | Description | Finding / Impact |
|---|---|---|
| github.com/garrytan/gbrain | Garry Tan's GBrain | Markdown + pgvector hybrid, compiled truth + timeline pages, dream cycles. High influence: adopted compiled truth pattern and dream cycle concept. adopted |
| github.com/letta-ai/letta | Letta (MemGPT) | OS-inspired tiered memory with agent paging between tiers. Sleep-time compute concept influenced vault tiered architecture. influenced |
| github.com/mem0ai/mem0 | Mem0 Production Memory | Universal memory layer (186M API calls/quarter). 26% accuracy improvement benchmark. Not adopted — API dependency reduces inspectability. reviewed |
| github.com/langchain-ai/langmem | LangChain Memory SDK | Semantic + episodic + procedural taxonomy. Good reference but too coupled to LangChain ecosystem. reviewed |
| github.com/microsoft/graphrag | Microsoft GraphRAG | LLM builds entity knowledge graphs. Decided against: graph database overhead not justified for single-user agent. reviewed |
| github.com/getzep/graphiti | Zep Graphiti | Temporal knowledge graphs, 94.8% DMR benchmark. Bi-temporal concept interesting but complex. reviewed |
| github.com/agiresearch/A-mem | A-MEM Agentic Memory | Zettelkasten-inspired memory with linked notes and memory evolution. High influence: new facts update old memories. SOTA on 6 models. adopted |
| github.com/modelcontextprotocol/.../memory | MCP Memory Server | Official MCP knowledge graph server. Reviewed as integration point; vault_mcp.py provides similar functionality natively. reviewed |
Theoretical grounding for memory architecture. 13 papers reviewed, 5 with high influence on design.
| URL | Description | Finding / Impact |
|---|---|---|
| arxiv.org/abs/2304.03442 | Park et al. (2023) — Generative Agents | Foundational. Memory stream + reflection + planning. "Reflection was single biggest contributor to believable behavior." Retrieval: recency × importance × relevance. foundational |
| arxiv.org/abs/2310.08560 | Packer et al. (2023) — MemGPT | OS-inspired tiered memory. Influenced vault architecture (working/short-term/long-term tiers). influenced |
| arxiv.org/abs/2305.10250 | Zhong et al. (2023) — MemoryBank | Ebbinghaus forgetting curves for AI memory. Influenced stale_after and memory health tracking. influenced |
| arxiv.org/abs/2308.15022 | Wu et al. (2023) — Recursive Summarizing for Dialogue | Hierarchical compression algorithm. Relevant for dream cycle "merge" operations. reviewed |
| arxiv.org/abs/2502.12110 | Xu et al. (2025) — A-MEM Agentic Memory | High influence. Zettelkasten for agents. Memory evolution: new facts trigger updates to old memories. SOTA on 6 models. adopted |
| arxiv.org/abs/2502.06975 | Pink et al. (2025) — Episodic Memory for Long-Term Agents | Tulving's episodic/semantic distinction. Consolidation pathway (episodic → semantic) = dream cycle concept. influenced |
| arxiv.org/abs/2504.13171 | Lin et al. (2025) — Sleep-time Compute | Key paper. Academic foundation for dream cycles. ~5x compute reduction, 13-18% accuracy improvement. From MemGPT founders. foundational |
| arxiv.org/abs/2504.19413 | Chhikara et al. (2025) — Mem0 Scalable Long-Term Memory | Production benchmark: 26% accuracy improvement, 91% lower latency, 90% cost savings. reviewed |
| arxiv.org/abs/2404.13501 | Zhang et al. (2024) — Memory Mechanism Survey | 39-page taxonomy. Three dimensions: sources, forms, operations. reviewed |
| arxiv.org/abs/2512.13564 | Hu et al. (2025) — Memory in the Age of AI Agents | Most comprehensive survey (47 authors). Broad landscape view. reviewed |
| arxiv.org/abs/2404.16130 | Edge et al. (2024) — Graph RAG: Local to Global | Microsoft's graph RAG paper. Reviewed but not adopted — too heavyweight for our scale. reviewed |
| arxiv.org/abs/2501.13956 | Rasmussen et al. (2025) — Zep Temporal Knowledge Graphs | Bi-temporal model. Good concept but complex implementation. reviewed |
| arxiv.org/abs/2501.01880 | Li et al. (2025) — Long Context vs. RAG | Validated our approach: at small corpus size, long context wins over RAG. FTS5 as hedge for growth. foundational |
Industry analysis that shaped architectural decisions. 13 posts from leading AI practitioners.
| URL | Description | Finding / Impact |
|---|---|---|
| anthropic.com/.../effective-context-engineering... | Anthropic Engineering | Core philosophy. "Smallest set of high-signal tokens." Keep memory lean. Directly influenced markdown-as-source-of-truth. foundational |
| mem0.ai/blog/state-of-ai-agent-memory-2026 | Mem0 Blog — State of AI Agent Memory 2026 | Comprehensive 2026 overview. Positioned PKA within the landscape. Validated file-based memory as mainstream. reviewed |
| letta.com/blog/agent-memory | Letta Blog — Agent Memory Deep Dive | Tiered memory concepts. Useful framing for working/short-term/long-term tiers. reviewed |
| huyenchip.com/2025/01/07/agents.html | Chip Huyen — "Agents" | Context Expansion Law: context expands to fill limits, so RAG stays relevant. Important contrarian view. influenced |
| hamel.dev/notes/llm/rag/not_dead.html | Hamel Husain — "Stop Saying RAG Is Dead" | RAG economics still matter. Graph DBs unnecessary. ColBERT > single-vector. reviewed |
| simonw.substack.com/.../things-we-learned... | Simon Willison — "Things We Learned About LLMs in 2024" | Long inputs > short prompts, but keep infra simple. Validated SQLite-everything philosophy. influenced |
| jxnl.co/.../facets-context-engineering/ | Jason Liu — "Beyond Chunks: Context Engineering" | Faceted search with structured extraction > top-k vector search. Influenced YAML frontmatter approach. influenced |
| sequoiacap.com/.../harrison-chase/ | Harrison Chase (LangChain) — Sequoia Podcast | Key quote: "If you're building a long-horizon agent, give it access to a file system." Validated file-based approach. foundational |
| cognition.ai/blog/devin-annual-performance-review-2025 | Cognition — Devin Annual Review 2025 | DeepWiki reference. Proprietary code-specialized indexing. Enterprise-grade but closed-source. reviewed |
| ragflow.io/blog/rag-review-2025-from-rag-to-context | RAGFlow — "2025: From RAG to Context" | RAG evolution overview. Context engineering superseding pure retrieval. reviewed |
| alexgarcia.xyz/.../sqlite-vec-hybrid-search/ | Alex Garcia — Hybrid Search with SQLite | FTS5 + sqlite-vec hybrid. Shows SQLite can do hybrid without Postgres. influenced |
| glaforge.dev/.../reciprocal-rank-fusion/ | Guillaume Laforge — Reciprocal Rank Fusion | RRF technical reference. Relevant if we add hybrid retrieval later. reviewed |
| dev.to/imaginex/.../markdown-files-all-you-need | ImagineX — "Markdown Files Are All You Need" | Community validation of our exact approach. Markdown files as complete memory solution. influenced |
Video content and community discussions that influenced design decisions.
| URL / Source | Description | Finding / Impact |
|---|---|---|
| youtu.be/mi4hcipESKQ | Nate Herk — Claude Code EA | Claude Code as Executive Assistant. Influenced assistant/dispatch model. influenced |
| Greg Isenberg / Vin (212K views) | Thinking Tools & Second Brain Philosophy | /ghost, /challenge, /trace, /connect, /ideas patterns. "Writing and daily reflection are the engine." Phase 2 thinking tools conceptualized. influenced |
| Mark Kashef (YouTube) | Obsidian Skills & Session Summary Workflows | Led to installing kepano/obsidian-skills. /tldr concept influenced session indexer. adopted |
| Brad (bradautomates, YouTube) | Brain Dump & Git Philosophy | Brain dump auto-classify, auto-commit hooks, grep-before-glob. Multiple patterns adopted into /brain skill. adopted |
| Cole (coleam00, YouTube) | Progressive Disclosure Pattern | "Metadata always in context, body when triggered, resources as needed." Adopted in skill architecture. adopted |
| Category | Count | Adopted/Influenced |
|---|---|---|
| Karpathy LLM Wiki | 2 URLs | 2 fully adopted |
| GitHub Templates | 5 URLs | 3 adopted |
| Memory Frameworks | 8 URLs | 3 influenced |
| Academic Papers | 13 URLs | 5 high-influence |
| Expert Commentary | 13 URLs | 4 directly shaped |
| YouTube / Community | 1 URL + 4 refs | 3 led to adoptions |
| TOTAL | 42 URLs + 4 refs | 20 adopted/influenced |
Generated by Luci | MC-602 | 2026-04-15