Afterhack
An open-source experimental project for AI agent memory systems that uses a graph database to implement claim provenance and error-correction propagation, emphasizing exact evidence matching and traversal of dependency chains.
Timeline (last 90 days)
In the Afterhack project, the extraction stage sets a hard rule: the model must return an evidence_quote for each claim, and the API verifies that the quote is an exact substring of the source document, requiring a character-by-character match.
On the agent side, Afterhack consists of two roles, Analyst and Coordinator, working in division of labor across two model calls: the Analyst has the ability to search claims and pull original snippets, and the Coordinator reviews and writes draft decisions.
In Afterhack, the Coordinator may only cite claims that were both searched and whose source snippets were pulled during the current run, and cannot fabricate IDs or reuse old IDs from memory.
If nothing survives filtering, the decision is marked needs_review with the title "No source-backed answer yet" and no action is produced.
The first version of Afterhack's correction code copied the evidence of the old claim directly onto the new claim, causing the replacement claim to point to the old evidence, with confidence set to 1.0.
The corrected error-correction workflow creates an independent fragment attributed to the Correction node and marked with a human source, replaces the claim to reference that fragment, and starts with confidence 0.0 while carrying confidence_kind.
In Afterhack, each workspace uses an independent graph whose name is derived from a server-generated UUID; requests cannot specify which graph to access, eliminating unauthorized access at the isolation level.
All Facts (7)
If nothing survives filtering, the decision is marked needs_review with the title "No source-backed answer yet" and no action is produced.
50%UnverifiedOn the agent side, Afterhack consists of two roles, Analyst and Coordinator, working in division of labor across two model calls: the Analyst has the ability to search claims and pull original snippets, and the Coordinator reviews and writes draft decisions.
50%UnverifiedIn Afterhack, the Coordinator may only cite claims that were both searched and whose source snippets were pulled during the current run, and cannot fabricate IDs or reuse old IDs from memory.
50%UnverifiedIn the Afterhack project, the extraction stage sets a hard rule: the model must return an evidence_quote for each claim, and the API verifies that the quote is an exact substring of the source document, requiring a character-by-character match.
50%UnverifiedIn Afterhack, each workspace uses an independent graph whose name is derived from a server-generated UUID; requests cannot specify which graph to access, eliminating unauthorized access at the isolation level.
50%UnverifiedThe corrected error-correction workflow creates an independent fragment attributed to the Correction node and marked with a human source, replaces the claim to reference that fragment, and starts with confidence 0.0 while carrying confidence_kind.
50%UnverifiedThe first version of Afterhack's correction code copied the evidence of the old claim directly onto the new claim, causing the replacement claim to point to the old evidence, with confidence set to 1.0.
50%