Summary
MemR3 (Dec 2025, arxiv 2512.20237) introduces evidence-gap tracking: during memory retrieval, explicitly track what evidence is still missing and do iterative retrieval rounds with reflection between them.
Currently, aletheia's recall pipeline does single-shot retrieval (vector search + Datalog query). The evidence-gap technique would make retrieval iterative:
- First retrieval round returns initial facts
- Reflection step identifies what's still unknown
- Second retrieval round targets the gaps
- Repeat until gaps are closed or budget exhausted
How this applies
The evidence-gap tracker is a form of retrieval self-audit — the system explicitly tracks "what do I still not know?" This could improve recall quality for complex queries that span multiple knowledge domains.
Source
MemR3: "Memory Retrieval via Reflective Reasoning" (arxiv 2512.20237, Dec 2025). Works with both chunk-based RAG and graph backends.
Summary
MemR3 (Dec 2025, arxiv 2512.20237) introduces evidence-gap tracking: during memory retrieval, explicitly track what evidence is still missing and do iterative retrieval rounds with reflection between them.
Currently, aletheia's recall pipeline does single-shot retrieval (vector search + Datalog query). The evidence-gap technique would make retrieval iterative:
How this applies
The evidence-gap tracker is a form of retrieval self-audit — the system explicitly tracks "what do I still not know?" This could improve recall quality for complex queries that span multiple knowledge domains.
Source
MemR3: "Memory Retrieval via Reflective Reasoning" (arxiv 2512.20237, Dec 2025). Works with both chunk-based RAG and graph backends.