Palimpsest

An original, Ethereum-style blockchain used as the temporal backbone of a semantic knowledge store — a proof of concept for agentic LLM frameworks that need to traverse meaning through time.

The problem

In a conventional embedding store, a topic like Pluto is one node carrying the compressed weight of its entire history: discovery, textbook planethood, doubt, demotion, redemption — all averaged into a single vector. Asking "what did Pluto mean in 1992?" isn't a lookup; it's archaeology. And asking "how did Pluto change between 1990 and 2010?" isn't answerable at all from one point in embedding space.

The idea

Give time its own dimension — and let a blockchain be that dimension. Every block is a temporal coordinate. A topic's semantic state (a real 768-dimensional embedding) is committed into contract storage block by block, so the chain's history is the topic's history:

Point-in-time — "Pluto at block 63 (1992)" is a native state lookup: storage_at(registry, key, 63). No decompression, no averaging.
Any window — "AI from 1990 to 2016" is a walk over block-height snapshots, reconstructed from write journals in microseconds.
Discontinuities — a reclassification, a genre pivot, a winter: each is a single-block semantic jump the chain can point to exactly.
Cross-topic, same-instant — compare two topics at the same block, something a flat store cannot even express.

The knowledge "explodes on arrival": an agent lands on a topic and the whole temporal structure is immediately traversable — any point, any window, any comparison — because the chain already indexed time for it. And nothing here is specific to time: block-ordered dimensions could equally model versions, jurisdictions, or viewpoints. Chains of different dimensions, one per axis of complexity.

What's actually running here

Chain

97 blocks

one block per year, 1930–2026, seeded & deterministic PoW-lite

Execution

256-bit VM

original ~40-opcode stack machine with gas, logs, revert

Semantics

768-dim real

nomic-embed vectors, int8-quantized into 24 storage words

Interpretability

16 axes = 1 word

named-axis projection packed into one 256-bit storage word

The demo: three timelines

Three popular topics from three domains, replayed onto the chain one block per year by an oracle, each era a short prose snapshot embedded with a real model:

Taylor Swift pop culture — continuous drift: country → pop → indie → imperial phase. Watch the genre_country and genre_pop axes cross in 2014.

Pluto pop science — the sharpest event in the dataset: the 2006 IAU demotion. The reclassification axis peaks exactly there — and the chain shows the semantic rupture actually began in 2000–2005, before the vote.

Artificial Intelligence pop history — boom and bust: the disillusion axis peaks at both AI winters (1973, 1990) and collapses in the deep-learning era.

Explore them live on the Timelines page, inspect raw chain state in the Explorer, step through contract execution in the Playground, or watch a block seal in Proof of Work.

Why a blockchain, really?

Three properties do the work. Ordered immutability: the block axis is a tamper-evident temporal index — every historical state carries a root hash the whole chain vouches for, so an agent can trust "Pluto-in-1992" as hard provenance, not a cache entry. Programmable writes: updates pass through contract logic (who may write, what shape, what events fire), so the knowledge store has enforceable semantics rather than conventions. Journaled state: every block records exactly what changed, which makes temporal queries — diffs, drift, discontinuity scans — cheap and exact rather than reconstructed after the fact.

Where this goes

The POC is single-node and one-dimensional (time). The obvious next steps: multiple topic registries sharded across chains; non-temporal dimensions (versions of a codebase, jurisdictions of a law, perspectives on an event) each as their own block-ordered axis; and agent-side retrieval that treats (topic × block-coordinate) as the addressing scheme for memory.