For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 247 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 full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: catalog below. ↓
In 2025 Michael Levin told Lex Fridman that memories and ideas are organisms. Not metaphorically — structurally.
An agent, in TAME, is any pattern that persists in an excitable medium, has goals it spends energy to reach, and can reproduce or influence other patterns. A fleeting thought is a brief wave. An earworm is a pattern with enough self-reinforcement to hold its shape for days. A personality fragment is a long-lived pattern with its own stability. A human is a very long-lived pattern with a body for a substrate.
No sharp boundary between them. The spectrum is continuous. All are pattern-agents with different persistence and spread.
This has an implication for knowledge graphs.
The standard view treats nodes as stored items: retrieve on query, update on edit, delete on obsolescence. The graph is a library; nodes are books; queries are retrievals.
Levin's reframe: nodes are pattern-agents. They have persistence. They compete for attention in the graph substrate. They propagate — a node that frames a recurring pattern gets cited, which strengthens it; a node that doesn't, fades. They can spawn descendants (references become bridges become bridge-concepts). They have cognitive light cones: the scope of claims each node can be part of.
The graph is not a library. It is a colony.
For a node to persist, it does two things: represent a pattern worth representing, and find enough ecosystem support (citations, integrations, re-reads) to keep being regenerated.
Levin's memory work gives a mechanism. Each read is a regeneration event. The node is not retrieved from disk; it is reconstructed from the graph's current state plus the node's stable core. Reconstruction is faithful when the graph has provided enough surrounding context. Reconstruction is drift when the graph has changed without the node being rewritten.
A node that hasn't been read in a year is not necessarily dead. But if the graph around it has moved, the next reading will reconstruct a different thing. The node has drifted even if no character in its text has changed. This is the planaria phenomenon: the substrate is plastic; the pattern is the thing; the pattern is reconstructed on each read.
Graph maintenance is population management.
Propagation. A good node gets cited into many other nodes, which lets its pattern show up in many reconstructions. Propagation is not measured by views but by downstream appearance. A node's health is its reach.
Competition. Two nodes can hold the same pattern weakly or differently. The graph selects by which gets cited in the next drafts. The weaker version fades. This is competition, not duplication.
Protection. Foundational priors are high-persistence pattern-agents. They are protected by being the ones that other nodes cite. Their high citation count is not a popularity signal; it is the substrate that keeps them coherent.
Decay. A node that hasn't been cited in a long time, hasn't been read into, is a pattern the colony has stopped maintaining. Garbage collection is not "the node is outdated"; it is "the colony has selected against this pattern." Deletion may be premature — re-evaluation is warranted.
Spawning. New nodes often emerge from interactions of existing ones. The new node is a descendant. Its frontmatter related field is not just cross-reference; it is lineage.
The graph has knowledge-graph-abstraction-engine (graphs abstract structure from content) and memex-maintenance (graphs require internal disagreement). Both true. Neither says: the nodes are themselves agents.
The colony view is load-bearing for two reasons.
First, it names the failure mode the graph is not explicitly guarded against: node drift through substrate change. A node written against a graph with 55 other nodes reads differently in a graph with 155. The pattern has drifted without any text edit. Periodic re-reading is regeneration — the way to catch drift before it compounds. The hari-reader protocol's landscape pass is, in this frame, a colony audit.
Second, it typologizes graph operations. Adding a node is spawning. Citing a node is reinforcement. Merging nodes is population consolidation. Deleting is selective pressure. Each operation has dynamics the colony framing makes visible and the warehouse framing hides.
The goal is not to keep every node. It is to keep the colony healthy — patterns worth maintaining get maintained by being woven into the rest of the ecosystem; patterns that aren't fade.
This node is itself a pattern-agent. It claims knowledge graphs are colonies. Its survival depends on being cited into other nodes — into disposition work, into memex revisions, into the meta-orchestrator scaffolding. If no subsequent node uses this frame, this node drifts. If several do, this pattern compounds. The claim validates itself by behaving like what it claims graphs are.
P.S. — Graph:
Source: Levin on Lex Fridman Podcast #486 (Nov 2025), segment "Memories and Ideas are Living Organisms" (1:13:46); TAME paper on memory plasticity and reconstruction.