Errata.
spend $——.———— · the homebuyer’s history
Memory that keeps its corrections — every answer cited, every revision kept, refusal when the history never says.

An agent-memory layer on HydraDB

Memory that keeps its corrections.

Errata stores what a conversation established as an append-only graph of claims. When a fact changes, nothing is overwritten: the new claim supersedes the old one, and both stay dated, cited, and queryable.

Watch it happen to one fact

The demo memory is one person's real chat history — 39 conversations · 396 turns · ingested read-only from LongMemEval. One thread runs through it: a mortgage pre-approval that would not sit still.

1 AUG 11 2023

“pre-approved for $350,000 from Wells Fargo.”

$350,000

2 NOV 30 2023

“I got pre-approved for $400,000 from Wells Fargo”

$400,000

3 AUG 18 2026 · filed through the API

“corrected by the user to $425,000”

$425,000

the chain as ingested — every strike is a SUPERSEDES edge; nothing was deleted, all of it is still queryable

“What was the amount I was pre-approved for when I got my mortgage from Wells Fargo?”

Ask the live graph

the answer arrives cited to the exact turn, with every displaced value struck through beside it — and a question the history never answers gets an honest refusal, not a guess

Measured, not asserted

On 150 LongMemEval questions, against the same reader model reading the entire history and against naive vector retrieval, Errata scores 60.0 overall to their 47.5 and 45.8 — at a fraction of the context — and the columns it loses are printed with the rest, every cell opening the judged rows behind it.

See the evidenceReplay the timeline