RETCON
An open-source memory management project for AI agents that records the provenance and dependency relationships of memories in a graph database. It supports cascading revocation of contaminated information and human-approved restoration.
Timeline (last 90 days)
RETCON is an open-source project that uses the graph database FalkorDB to record the justification of agent memories (sources, duplicates, plan dependencies) to support cascading retraction.
RETCON's core rule is: as long as at least one justification of a memory remains intact, it is believed; on retraction its state is marked OUT rather than deleted, and it can be restored.
In the RETCON demo, a pending payment of $34,500 was blocked, and the $6,000 already paid entered a recall list with a drafted bank recall request, but it will not be sent before human approval.
RETCON provides three action buttons: Preview (run the retraction on a copy of the graph and show the impact), Commit (execute the real retraction) and Reinstate (restore).
The RETCON system sets a gating rule: any payment over $10,000 that depends on fewer than two independent sources is rejected.
In the simple scenario of the benchmark (1,024 memories), the three rules of deleting the source, deleting similar content, and invalidating one hop left 87% to 94% contamination residue.
In the exact scenario of the benchmark, the "contaminate all downstream" rule falsely flagged 36 honest payments (totaling $1,163,150, of which 18 had been paid), while RETCON flagged 0.
In 40 real retractions replayed from the Retraction Watch database, the "contaminate all downstream" rule lost 24 topic-level stances supported by multiple independent papers, while RETCON lost 0.
Truth maintenance systems (TMS) were proposed in the late 1970s, with the core idea of recording a justification for each belief and recursively retracting beliefs when justifications fail; RETCON extends this to a cross-agent, cross-session distributed memory layer.
All Facts (9)
Truth maintenance systems (TMS) were proposed in the late 1970s, with the core idea of recording a justification for each belief and recursively retracting beliefs when justifications fail; RETCON extends this to a cross-agent, cross-session distributed memory layer.
50%UnverifiedRETCON's core rule is: as long as at least one justification of a memory remains intact, it is believed; on retraction its state is marked OUT rather than deleted, and it can be restored.
50%UnverifiedRETCON is an open-source project that uses the graph database FalkorDB to record the justification of agent memories (sources, duplicates, plan dependencies) to support cascading retraction.
50%UnverifiedRETCON provides three action buttons: Preview (run the retraction on a copy of the graph and show the impact), Commit (execute the real retraction) and Reinstate (restore).
50%UnverifiedThe RETCON system sets a gating rule: any payment over $10,000 that depends on fewer than two independent sources is rejected.
50%UnverifiedIn the simple scenario of the benchmark (1,024 memories), the three rules of deleting the source, deleting similar content, and invalidating one hop left 87% to 94% contamination residue.
50%UnverifiedIn the exact scenario of the benchmark, the "contaminate all downstream" rule falsely flagged 36 honest payments (totaling $1,163,150, of which 18 had been paid), while RETCON flagged 0.
50%UnverifiedIn 40 real retractions replayed from the Retraction Watch database, the "contaminate all downstream" rule lost 24 topic-level stances supported by multiple independent papers, while RETCON lost 0.
50%UnverifiedIn the RETCON demo, a pending payment of $34,500 was blocked, and the $6,000 already paid entered a recall list with a drafted bank recall request, but it will not be sent before human approval.
50%