Home / Work / Debias Agent memory · Graph Hacks 2026

Debias

A bias-aware memory graph for AI agents, built on FalkorDB for Graph Hacks (WeMakeDevs × FalkorDB), October 2026.

The problem

Agents remember, but their memories have no immune system. A user's offhand claim gets stored as fact. A conclusion the agent invented three sessions ago gets cited as evidence today. A stale note recalled four times starts to feel true because it's familiar.

The approach

Debias writes every memory with its provenance. Before the agent acts, five pure-Cypher detectors check the claims it's about to rely on. No LLM judges anything: the graph itself shows why a memory shouldn't be trusted, and the agent either gets a warning in its prompt or escalates to a human.

with debias("support") as guard:
    check = guard.relies_on([claim])
    if check.flags:
        system_prompt += "\n" + check.note

View on GitHub →

Want this kind of testing on your system? [email protected]