Why Venture Capital Is Betting on AI Digital Brains

In April, Andrej Karpathy posted a GitHub gist describing how he stopped spending most of his AI tokens on code and started spending them on a wiki an agent writes and maintains. Within weeks, investors gave the pattern a name – the digital brain – and it now sits atop a funded infrastructure category. The tools cost next to nothing, yet the moat belongs to whoever starts compiling first.

A digital brain is a persistent knowledge base an AI agent builds, cross-links and corrects over time. It is not a pile of documents the model re-reads on every question. Karpathy's idea file specifies three layers: immutable raw sources, wiki pages the agent writes, and a schema file that tells the agent how to behave. Three operations run the system: ingest a source, query the wiki, lint for contradictions and orphaned pages. The contrast with retrieval-augmented generation is accumulation. RAG rediscovers knowledge from scratch each query. The wiki compounds.

Venture capital has a specific reason to care. Crunchbase counted $300 billion of global venture investment in the first quarter of 2026, with four AI rounds accounting for 65% of it. Firms deploying that capital are themselves the slowest-digitized professional service in the stack, and the gap is closing fast. Affinity's survey of nearly 300 private capital professionals found 85% automating daily tasks with AI and 82% using it for sourcing research, up from 76% and 64% a year earlier. Stephen Lantz, a principal at Bain Capital, told Affinity he pushes raw meeting notes through Claude and has the output posted back to the firm's CRM with entity tags attached. That loop, notes in, structured memory out, is a digital brain with a vendor database instead of a folder of Markdown.

The labs are building the plumbing. Anthropic's finance connectors wire Claude into PitchBook, FactSet, S&P Capital IQ, Morningstar and firms' own CRMs and research repositories under governed access. The same company closed a $65 billion Series H in May at a $965 billion valuation, with run-rate revenue past $47 billion. Goldman Sachs, Blackstone and Hellman & Friedman stood up a $1.5 billion deployment venture to embed Claude inside portfolio companies. At the far end, a pre-launch fund calling itself Claude VC says agent teams will run sourcing, diligence, memo drafting and LP qualification. Whether a fund with no human judgment in the loop can raise from institutional LPs is an open question.

Below the firms sits a memory layer investors have already recognized, funded and priced. Mem0 raised $24 million across seed and Series A, led by Kindred Ventures and Basis Set Ventures with Peak XV, GitHub Fund and Y Combinator. Its API calls rose from 35 million in the first quarter of 2025 to 186 million in the third. Supermemory, founded by a 19-year-old who briefly worked at Mem0, closed a $2.6 million seed from Susa Ventures, Browder Capital and SF1.vc with Google's Jeff Dean and Cloudflare CTO Dane Knecht as angels. It started life as a consumer second brain app before pivoting to an API. Letta, backed by Felicis, sells a full stateful agent runtime. Zep sells a temporal knowledge graph.

The benchmark claims deserve skepticism. A Q3 2026 comparison by Mnemoverse notes that Mem0's published LongMemEval score fell to 73.8% when reproduced under a rival vendor's harness, and that no published number survives without knowing who ran the test. Three memory startups together raised less than a tenth of one percent of what Anthropic raised in a single round, and every frontier lab now ships native memory. A standalone memory API is a feature labs can absorb. The curated corpus a firm builds on top of it is not.

Goldman Sachs (GS), which participated in the $1.5 billion deployment venture, carries an Alpha Score of 58/100, indicating moderate momentum. GS stock page

Building a digital brain today takes an afternoon and no budget beyond model tokens. Install Obsidian, free for personal use, and create a vault with three folders: raw, wiki, and a single schema file named CLAUDE.md or AGENTS.md. Open Claude Code or Codex in that folder and paste Karpathy's gist as the first instruction. Drop ten sources into raw – meeting notes, deal memos, clipped articles, a few PDFs – and ask the agent to ingest them. Ask a question that requires synthesizing three of them. Run a lint pass. Commit the folder to git. The open-source claude-obsidian plugin packages these steps into slash commands for anyone who wants less setup. Ten sources is enough to learn whether the ingest, query and lint loop fits how you work. A hundred is where it starts to outperform search.

For investors, the question is where portfolio memory lives. Firms that let it accumulate inside a vendor's database are renting their own institutional knowledge. Firms that compile it into files they own can swap models, swap memory APIs, and keep the asset. For founders selling memory infrastructure, the window is narrowing: persistence is becoming table stakes rather than a product. For everyone else, the cost of not building is the one Karpathy describes: an agent that rediscovers your world from scratch every morning, while the competitor next door started compiling yesterday.