Overview
An agent that remembers is only useful if its memory stays honest. Tell it in January that you live in Sydney and in March that you moved to Melbourne, and when someone later asks where you live, something has to decide which stored fact still holds and which has been overtaken.
This is where my honours research lives: conflict resolution for agent memory, studied through the lens of question answering. When the facts an agent has saved contradict one another, answering correctly means finding the conflicting set, working out which version is current, and reasoning over it before it speaks. The work spans two settings: single-hop questions that turn on one fact, and multi-hop questions that chain several facts together, where a single stale link breaks the whole answer.
How the study is built
- A baseline picture first. I evaluate a range of methods on both splits, from embedding-based retrieval that ranks stored facts by relevance (including a time-decay re-ranking variant that favours the recent) through to RAG and memory agents that organise, update, and recall facts on their own.
- From those baselines, a failure taxonomy: a map of where and why conflict resolution breaks down, so the fixes target real failure modes rather than guesses.
- Then the improvements: detecting and grouping conflicting facts and reasoning over them before answering. The final design is settled through experiment, not decided up front.
- For the harder multi-hop split, query parsing: breaking a compound question into its underlying sub-questions, an ordered chain or a tree, so each hop resolves against the right set of conflicting facts.
What it contributes
An evaluation and failure analysis of how these methods handle contradiction, and a pipeline that improves conflict-resolution accuracy, with the multi-hop split (FC-MH) as the main target since it is the more demanding and less studied case. The output is a research report on the approach, baselines, results, and limitations, backed by a working implementation.
What I'm learning
That most of the hard part isn't storing memories. It's deciding, gracefully, which version of the truth to keep. It turns out remembering well and forgetting well are the same skill.