Idea Lab · Oct 16 2026 · Fleet winner · Priority task
Temporal Entanglement Mapping (TEM)
A design for agent memory that survives a server dying. Every memory fragment is signed from its content and context, linked to the fragments it was derived from (a causal graph), and copied to other nodes. When a node fails, the survivors rebuild what they can from those copies and check every copy against its signature.
Toy simulation · example nodes and memories
Five nodes hold 24 memory fragments. Click a node to kill or revive it, then press Recover. Squares are fragments: their colour shows whether they survived.
Intact on a live homeRecovered from a copySurvived, but a parent is lost (reasoning gap)LostNot yet checked
–Recovered from copies
–Lost for good
–Reasoning gaps
–Slowest recovery fetch (network only)
Claimed by the proposal, not verified
"Recovery times under 20ms regardless of geographic dispersion." A round trip between the US East Coast and Southeast Asia is well over 150 ms on today's internet, so any recovery that has to fetch across that distance cannot finish in 20 ms. The round-trip times on this page are rough illustrative figures.
"Surviving peers traverse the graph to reconstruct missing reasoning chains." A graph records which fragments came from which. It cannot recreate a fragment whose only copies were on the dead node; the toy shows those as lost.
"Inherent resilience against Byzantine faults through quorum-based semantic verification." Signatures catch corrupted copies. Lying nodes need a real quorum protocol, which is not specified.
"Episodic Tensor Store" with "differentiable tensors" has no clear role in persistence and is left out here.
What to try
With 2 copies, kill Node D and recover. Everything comes back. Now kill C and D together: fragments whose only two copies were on C and D are lost, and their children show as reasoning gaps.
Set copies to 1. Any dead node means lost memory, the single-point-of-failure case TEM wants to remove.
Set copies to 3, press Corrupt a stored copy, kill that fragment's home node, and recover. The bad copy fails its signature check and the next holder is used.
Kill Node A (US-East) and look at the slowest fetch time when the copy has to come from Asia.
Press Partition, write, heal. Both sides write to the same key while split; on heal, the merge keeps both and flags the conflict instead of silently picking one.
A real build for this fleet could be much smaller than TEM: an append-only memory log per agent, each entry hashed and listing its parent entries, replicated to two other VPS nodes, with a nightly check that every copy still matches its hash.