Nate's core claim across this cluster: memory, not intelligence, is the binding constraint on AI usefulness, and it's getting relatively worse over time — intelligence has scaled roughly 60,000x since ChatGPT launched while memory infrastructure has scaled maybe 100x. Every new chat starts from zero; every tool switch discards accumulated context; vendors keep selling bigger context windows as the fix when the real problem is architecture, not volume.
He starts from a technical debunking. Long context windows are marketing more than engineering: a "200K token" model behaves like it has ~20K of reliable memory because attention is O(n²), position-biased toward the start and end of a prompt, and under-trained on far-distant tokens. The "lost in the middle" effect is real and reproducible — feeding a model more documents can make it perform worse, not better. His prescribed workaround, still valid throughout the cluster: don't trust the context window, architect around it with chunking, summary chains, retrieval (RAG), and explicit context budgeting by task type (planning needs breadth, execution needs precision).
From there the cluster tracks a genuine evolution in his prescribed system, in three distinct eras:
Era 1 — "Second Brain" (mid-2025 into early 2026). Pick a human-facing tool (Notion, Obsidian, or Mem — no universally correct choice, match to your style and ADHD-or-not tolerance for structure), run a 4-week capture habit, then layer AI summarize/extract/connect on top. By January 2026 this hardened into an explicit no-code architecture: Slack (or equivalent) as a frictionless single drop box, an AI "sorter" that classifies without asking you to tag anything, a Notion filing cabinet, a logged "receipt" for every action, a confidence "bouncer" that routes uncertain items to human review instead of guessing, and a daily/weekly "tap on the shoulder" digest. Twelve build principles came with it — reduce the human's job to one behavior, separate memory/compute/interface, treat prompts as APIs not creative writing, design for restart not perfection.
Era 2 — "Open Brain" (March 2026, the pivot). Nate explicitly retired the Second Brain guide: it was built for humans reading Notion pages, not for agents that need to query your knowledge directly. Open Brain replaces the Notion/Zapier middleman stack with a Postgres database plus pgvector embeddings, exposed over MCP so any AI client (Claude, ChatGPT, Cursor, whatever ships next) reads and writes the same store — no per-tool silos, no vendor lock-in, ~$0.10–0.30/month to run. The architecture is explicitly agent-first: agent-readable and agent-writable, human-readable and human-writable, on the same tables.
Era 3 — extensions and "Open Stack" (March–July 2026). Once Open Brain existed, the question became what to build on it. Key additions: a "two-door" principle (every extension needs an agent-facing query path and a human-facing visual one, because pipelines and calendars are inherently spatial, not conversational); a hard "judgment line" — agents surface, humans decide, agents execute, never blurred; and a write-time-vs-query-time framing borrowed from a viral Karpathy post (a compiled wiki front-loads synthesis at ingest time; Open Brain does synthesis fresh at query time) resolved into a hybrid: database as source of truth, a scheduled "wiki compiler" recipe as the browsable, regenerable view on top. By mid-2026 Nate reframed the whole endeavor as three owned layers — Open Brain (memory), Open Skills (method/how-to), Open Engine (task/handoff state) — sitting underneath whatever "rented" frontier model you plug in that week, because the model is disposable but your accumulated context is not.
A parallel thread throughout: memory is portable professional capital you don't currently own. Six months of real AI use encodes domain knowledge, workflow calibration, and a behavioral relationship the tool has learned about you — and all of it evaporates at a platform switch, an enterprise IT migration, or a job change. Nate's answer is the same architecture bet: own the database, speak the open protocol (MCP), keep your context extractable as a portable bundle instead of trapped in vendor "memory" features.
Key moves
Don't trust advertised context windows — budget for effective context (~20-50% of the sticker number) and architect with chunking, RAG, and summary chains instead
Treat memory as multiple separate problems (preferences, facts, knowledge, episodic, procedural) — each needs its own storage and retrieval pattern, not one undifferentiated blob
Separate context by lifecycle: permanent preferences ≠ project facts ≠ session state; mixing them is what breaks memory systems
Match storage to query pattern — key-value for preferences, SQL for exact facts, vectors for semantic recall, event logs for "what did we try before"
Compression is curation, not automation: write the brief yourself (or with AI amplifying your judgment); don't upload raw documents and hope the model extracts what matters
Build a frictionless single capture point ("drop box") — one inbox, zero decisions at capture time; multiple capture options is the friction that kills systems
Add a confidence "bouncer": below-threshold classifications should ask for clarification, not guess and pollute the database
Log every automated action as a "receipt" and make corrections trivial (a one-line "fix:" reply) — trust, not cleverness, is what keeps a system alive
Design for restart, not perfection — missing two weeks should cost a 10-minute catch-up, not a guilt spiral that kills adoption
Prefer agent-native, protocol-based infrastructure (Postgres + MCP) over app-native "second brain" tools once agents — not just you — need to query the data
Use NotebookLM (or similar) for accurate retrieval and a separate reasoning LLM (Claude/ChatGPT) for synthesis — retrieval and thinking are architecturally different jobs, don't force one tool to do both
Draw a hard line between agent and human labor: the agent surfaces patterns and holds the timeline; only the human decides and approves anything irreversible
Keep your AI working context portable — extract it into a model-agnostic bundle/profile you carry between tools, jobs, and employers rather than letting it accumulate inside a vendor's proprietary memory feature
When choosing between a compiled wiki and a queryable database, let the database stay authoritative and regenerate the wiki from it — never let synthesis drift become the source of truth
2026-03-13 — You built an AI memory system. Now your agent needs hands. — Introduces the "two-door" (agent + human interface) principle and the judgment line: agents surface and bridge time across data, humans decide, agents execute.
2026-03-02 — Why your AI starts from zero every time you open a new chat — The pivot post: retires the Notion/Zapier "Second Brain" in favor of "Open Brain," a Postgres + MCP system any AI client can query directly, ending platform memory silos.
2025-10-29 — I Spent 3 Months Testing Knowledge Systems for AI — Recommends NotebookLM for low-hallucination retrieval, paired with a separate reasoning LLM for synthesis — a deliberate two-step retrieval-then-thinking workflow.
2025-10-16 — I Wrote the AI Memory Fix Every Existing Solution Missed — Lays out 8 durable architecture principles (memory as architecture not feature, separate by lifecycle, match storage to query pattern, portability, compression as curation, verification, structured compounding).
2025-07-07 — Context Windows Are a Lie: The Myth Blocking AGI — Technical takedown of advertised context windows; "lost in the middle" is real, and the fix is chunking, retrieval, and context budgeting, not waiting for bigger windows.
2024-12-24 — The Half Trillion Dollar Memory Problem — Earliest post in the cluster: does the math showing true long-term AI memory at consumer scale would cost hundreds of billions in storage alone, framing memory as the next real AI bottleneck.