For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 780 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
The infrastructure to give an AI a version-controlled mind is half-built, and the missing half is the interesting one. The market has converged, fast, on how to keep many agents from stepping on each other. It has barely started on how to decide what they collectively know.
Look at what does exist, because it is real and maturing. Around parallel agents, git worktrees became the consensus isolation primitive in about eighteen months: Conductor runs parallel Claude Code and Codex sessions in worktrees, Sculptor runs them in Docker containers, and a crowd of others — Claude Squad, Superset, Switchboard, Crystal before it folded into Nimbalyst — give you a dashboard, a diff, and a merge button per agent. Around memory, an entire product category materialized in about two years: Mem0 bolts a vector-and-graph memory onto any agent, Zep keeps a temporal knowledge graph of facts that change over time, Letta pages an agent's memory in and out like an operating system. Around data, the git-for-datasets market kept consolidating, with lakeFS taking over stewardship of DVC, and Dolt and Pachyderm holding the database and pipeline corners. And a small front of git-native agent projects treats the repository itself as the agent: GitAgent and Memoir keep identity, rules, memory, and skills as version-controlled files, while Cloudflare's Artifacts offers a versioned, Git-compatible filesystem beneath them.
So three layers are built: isolation, so agents work without colliding; memory, so a single agent persists and recalls; storage, so any of it can be versioned. What none of them builds is the layer that decides, when two copies of an evolving mind disagree, which knowledge is true. The orchestrators come closest and then stop exactly at the threshold. They isolate the work and they assist the merge, but the merge decision lands on a human, and the whole stack is tuned for code.
Code is why the gap is invisible until you hit it, because code comes with an oracle. When two branches diverge, a test suite can tell you which merge is correct, or at least which one is broken. The agreement question has an external referee. Knowledge has none. When two copies of a mind hold different notes, different priors, different half-finished essays, there is no test that returns green for "true." This is not a small engineering miss. It is the reason the mature toolchain dead-ends at the human: the part that automated cleanly was the part with an oracle, and thought does not have one.
The research front already knows this, which is the surest sign it is the real frontier. The 2026 multi-agent memory surveys say it plainly: without coordination, agents overwrite each other, read stale state, and build on inconsistent versions of shared facts, and a last-writer-wins policy is unsafe because a weaker agent can clobber a better assessment. The proposed fixes all point the same direction — track provenance, author, confidence, and evidence on every memory; let an orchestrator or a trust model arbitrate; teach the system to tell a real conflict from a harmless overlap. There is now a named dataset of merge conflicts in AI-agent pull requests. The names for the missing layer exist in papers before they exist in products.
What the missing half does is reconcile: reconstruct the union of what every copy uniquely holds, weight conflicting claims by provenance and confidence rather than by recency or authority, merge knowledge semantically instead of line by line, and fail loud when it cannot. That is the AI GitHub in the sense that matters. It is not versioned storage for one agent's notes, which we have; it is a truth-establishment layer for many divergent copies of a thinking thing, which we do not.
It is half-built for a defensible reason, and that reason also says who builds the rest. The half that shipped is the half that generalizes across every codebase, because a test suite is a universal oracle. The half that hasn't is the half where the oracle is specific to the mind being merged — its priors, its taste, its provenance, what it means by true. That layer cannot be a thin product bolted onto any agent. It has to know the mind it reconciles. The first working version of it will not look like a platform. It will look like one mind that learned to keep its own copies honest, written down where the next one can read it.