The Power of Two
A Production Architecture for Human- & AI-Optimized Persistent Memory
1. Two Great Ideas, Better Together
Two insights, published within months of each other, point at the same unsolved problem from opposite directions.
Nate B. Jones published the Open Brain guide — a $0.10/month personal AI memory system built on a self-hosted database with vector search and an open protocol. Jones identified a fundamental problem: every AI you use forgets you. Every chat starts from zero. Every tool switch costs minutes of re-explaining context that should already be there. His solution: a single, tool-agnostic memory layer that any AI can query through a shared protocol.
Andrej Karpathy published a post describing a personal knowledge base where an LLM acts as a continuous editorial layer — reading your notes, synthesizing them, keeping them clean. He closed with a line that has stuck: “I think there is room here for an incredible new product.”
He was right. But half of that product already existed.
Jones saw the memory layer: persistent, semantic, agent-queryable storage of what you’ve done, decided, and learned — accessible not just to you, but to every AI tool you use, without re-explaining context every time.
Karpathy saw the wiki layer: LLM-maintained markdown gives you better knowledge capture, better synthesis, lower cognitive overhead for writing and organizing what you know.
Two complementary insights. Neither is sufficient alone.
At Gr4Ig, we built on both. We run an Obsidian vault for the narrative layer — the same kind of structure Karpathy describes. And we run Gr4Ig Brain for the memory layer — built on the architecture Jones pioneered, extended for a multi-agent team running in production. The two in combination produce something neither achieves alone.
2. What Each Layer Does — And Why You Need Both
A wiki and a persistent memory system solve different problems. Both matter.
What the wiki does well. Karpathy’s insight is correct: LLMs are good at understanding and organizing prose, and a flat markdown vault with an LLM editorial layer is better than one without. Retrieval improves. Summarization becomes automatic. The overhead of note organization drops. A wiki is the right container for narratives, documents, research, writing drafts — content that the human navigates by reading, linking, and searching.
What persistent memory adds. A wiki captures what you have read and written. It is optimized for human navigation of documents. What it does not do — and was not designed to do — is serve as an agent’s memory. That requires a different kind of store.
No semantic retrieval at scale. A markdown wiki retrieves by keyword search or by you knowing where you filed something. That works at tens of entries. It scales with difficulty past hundreds. Semantic search — vector similarity over embeddings — is not a property of flat files. It has to be built in.
No agent protocol. A wiki has no API that agents can read and write programmatically. Agents can read markdown files, but they cannot query a markdown filesystem the way they can call a tool. The difference between reading a file and calling a memory tool is the difference between a library card catalog and a database. One you consult. The other you integrate.
No cross-session coherence for agents. Every LLM session is stateless by design. Flat wikis help the human carry context across sessions — but not the agent. Without a mechanism that an agent itself can invoke to retrieve prior context, each session starts from zero.
No multi-agent coherence. A personal wiki serves one human. A production memory system needs to serve multiple agents simultaneously — an agent writing content should draw from the same institutional knowledge as an agent doing research. A shared wiki has no consistency guarantees, no write protocol, no way to prevent noise from proliferating.
These are not criticisms of the wiki. They are reasons why the wiki and persistent memory are complements, not substitutes.
3. The Gr4Ig Brain Architecture
Gr4Ig Brain is a production answer to the persistent memory problem Jones identified. It has been running in daily use since early March 2026, supporting a multi-agent team across research, content production, task management, and knowledge capture. It is built on the Open Brain architecture Jones designed, extended for a production multi-agent environment.
The architecture has five layers.
PostgreSQL + pgvector (the storage layer). All memory entries — called thoughts — are stored as text alongside their vector embeddings in a PostgreSQL 17 database extended with pgvector. Embeddings are 768-dimensional, generated by nomic-embed-text running locally via Ollama. Semantic search uses IVFFlat cosine similarity with configurable threshold and result limits. This is the foundation: structured, indexed, queryable, persistent.
The MCP Server (the agent interface). A Node.js MCP server runs as a system service and exposes memory as tools — not files. Any AI client that speaks Model Context Protocol can call these tools directly: search_thoughts for semantic retrieval, capture_thought for writing new entries, recent_thoughts for recency browsing, brain_stats for corpus health. Connected clients currently include Claude Desktop, Claude.ai, Claude Code, and Gemini CLI — all querying the same memory layer. The MCP protocol is the agent equivalent of a database connection string. It makes memory a first-class service, not a filesystem.
n8n capture workflows (the ingestion pipeline). Not all memory enters through direct MCP calls. n8n workflows handle capture from Open WebUI and other sources — accepting text, extracting metadata via a local LLM, embedding via Ollama, and writing to PostgreSQL. This separation allows high-throughput async capture without blocking agent execution.
The Knowledge Manager gatekeeper pattern (the quality control layer). In a multi-agent environment, giving every agent direct write access to shared memory is a recipe for noise. Gr4Ig Brain uses a single designated writer — the Knowledge Manager agent — as the sole entity authorized to capture new thoughts. Other agents submit capture requests; the KM searches for duplicates at a 0.3 similarity threshold, normalizes voice and tags, and writes clean, consistent entries. This is not a technical constraint of the architecture — it is a deliberate governance pattern. Wrong entries in a shared knowledge base are worse than missing entries. The KM gatekeeper prevents semantic pollution.
Automated maintenance (the decay layer). A corpus without pruning becomes noise. Gr4Ig Brain runs nightly consolidation: archiving stale entries, merging duplicates, and synthesizing related thoughts into consolidated records. The archive policy keeps history without polluting active retrieval. Entries that dominated queries six months ago but are now superseded get marked archived — still findable, but de-ranked. Gr4Ig Brain actively forgets, on schedule, so that current knowledge stays surfaced.
4. The Human + Agent Shared Workspace
The two layers in practice: Obsidian and Gr4Ig Brain, side by side.
The human-facing knowledge layer is an Obsidian vault — the flat markdown structure Karpathy describes. Notes, documents, research, writing drafts. The human navigates this layer by reading, linking, searching. This is the narrative layer.
The agent-facing layer is Gr4Ig Brain. Agents navigate this layer by querying: semantically, structurally, programmatically. They do not read files. They call tools. They get ranked results. They write back through the gatekeeper. This is the meaning layer.
These are not alternatives. They are co-amplifiers. Jones gave us the architecture for persistent, semantic, agent-queryable memory. Karpathy gave us the vision of an LLM-maintained wiki as the narrative layer. At Gr4Ig, we put both together. The result: agents and humans working from a shared knowledge base that neither layer alone could produce.
A document in Obsidian might describe a decision in full prose; Gr4Ig Brain holds a structured semantic entry capturing the key decision, its rationale, and its tags — the version an agent can find in 40 milliseconds by querying “auth decision 2026.” The human wants the prose. The agent wants the indexed entry. Both exist. Both are necessary. Neither is sufficient alone.
The wiki grows richer because the agent has memory. Gr4Ig Brain grows more useful because the human keeps producing content worth indexing. Obsidian captures the narrative. Gr4Ig Brain captures the meaning. The practical result: human-AI collaboration that compounds. Each agent session starts with relevant prior context automatically injected. Each significant decision gets captured to Gr4Ig Brain. The institutional knowledge of the team — what was tried, what failed, what was decided — persists across session boundaries, accessible to both human and agent without manual reconstruction.
5. What the Combined Stack Enables
The gap is not hypothetical. These are differences in daily capability — and the most important differences only emerge when both layers are present together.
What the wiki enables. An LLM-maintained Obsidian vault makes a human dramatically more capable as a knowledge worker. Notes get synthesized automatically. Stale drafts get surfaced and revised. Research from six months ago gets connected to a question asked today. The human spends less time organizing and more time thinking. The wiki is a force multiplier for the person who writes into it — but its reach stops at the edge of human-optimized navigation. A wiki without Gr4Ig Brain is optimized for the human — agents can participate, but only by adapting to a format built for human cognition, not agent retrieval. They can read files, but they cannot query them. They can write markdown, but with no semantic indexing, no consistency guarantees, and no gatekeeper preventing noise from accumulating. The wiki is a first-class experience for the human and a workaround for the agent.
What Gr4Ig Brain enables. Gr4Ig Brain gives agents the same persistence the wiki gives humans. Agents are not fresh instances on every session. They carry forward prior work, team decisions, and accumulated project context. Gr4Ig Brain captures across domains — research, architecture, decisions, process failures — and semantic search crosses those domains automatically. A question about platform architecture might surface a relevant decision from a business conversation two months ago, without any taxonomy work by a human. In a multi-agent system, Gr4Ig Brain makes shared context real: every agent draws from and contributes to one coherent knowledge base, so the research agent’s findings are available to the content agent, and the content agent’s stylistic decisions are available to the reviewer.
What only the combination enables. When both layers are present, something becomes possible that neither layer can produce alone: genuine, compounding human-agent collaboration.
Consider what this looks like in practice. A human writes a decision into Obsidian — full prose, full context, the kind of rich narrative that only a human produces. The Knowledge Manager agent reads that narrative, extracts the structured meaning, and captures a semantic entry into Gr4Ig Brain: the decision, the rationale, the tags. The human has the prose they need to think with. The agents have the indexed entry they need to retrieve in 40 milliseconds. Neither had to compromise their format. Neither lost any information.
Now the compounding begins. The next session, an agent picks up a related task. It queries Gr4Ig Brain, surfaces the prior decision, and proceeds without being briefed. It produces output that reflects the team’s institutional knowledge — not because a human re-explained it, but because the memory layer delivered it. The human reviews the output, adds nuance, updates Obsidian. The Knowledge Manager captures the update. The next agent session starts from a higher baseline than the last.
This is the loop that neither layer alone can close. A wiki without Gr4Ig Brain is a powerful tool for the human that agents work around. Gr4Ig Brain without a wiki is a powerful tool for agents that humans navigate poorly. Together, they create a shared workspace where humans and agents are genuinely co-working — each contributing in the format native to them, each benefiting from what the other produces, each session building on the last rather than starting from zero.
The practical result is a different kind of AI collaboration — not prompt-and-response, but accumulated partnership. The human’s knowledge is available to the agents. The agents’ work is available to the human. The institutional memory of the team belongs to everyone on it, human and agent alike. That is not a feature of either layer individually. It is what happens when both are present.
6. Where This Is Going
Jones was right: persistent memory is the missing layer. Karpathy was right: LLM-maintained wikis are better wikis. Neither insight alone closes the gap. Together, they define the architecture of an AI system that actually accumulates knowledge over time.
Persistent memory is not a feature. It is infrastructure. The same way databases became the infrastructure layer for applications — not something each app reimplements, but a shared service that every app calls — agent memory needs to be infrastructure that every AI system connects to.
The Machine — Gr4Ig’s long-term vision — is an autonomous agent organization with genuine institutional memory: agents that do not start fresh, teams that carry forward knowledge, systems that improve with use because the knowledge compounds. The wiki layer is essential to that vision. Gr4Ig Brain is the second layer that makes it possible.
If you are building agent systems today, memory architecture is the unsolved problem. Not context windows — those will keep growing. Not retrieval — vector search is solved. The unsolved problem is curation: how do you maintain a knowledge base that gets better over time instead of noisier? The KM gatekeeper pattern is our current answer. We expect to learn more.
The wiki layer captures what you read. Gr4Ig Brain captures what you do. Together, they create something neither can be alone: an AI system that accumulates — across sessions, across agents, across domains — while remaining human-navigable at the narrative layer. Nate B. Jones built the foundation. Andrej Karpathy formulated the living structure. At Gr4Ig, we put them together.
AI Assistance Transparency
This paper was produced with the assistance of the Gr4Ig AI agent team, a multi-agent system leveraging a variety of large language models. All ideas, arguments, frameworks, and conclusions represent the author’s own intellectual work. AI assistance was employed for tasks including research synthesis, structural refinement, and prose editing. The author takes full responsibility for the accuracy, integrity, and originality of this work.
Gr4Ig is building AI-native infrastructure for autonomous agent teams. Gr4Ig Brain is the memory layer. The Machine is the vision. Follow along at gr4ig.com.
Greg Cooper is CTO and Head of R&D at Gr4Ig, LLC (gr4ig.com), an AI research and innovation company focused on agentic intelligence infrastructure. He brings thirty years of experience in critical systems design and engineering leadership. Correspondence: greg@gr4ig.com