Infini Memory: Treating Agent Memory as Maintainable Topic Documents

arXiv·low signalXBlueskyLinkedInCopy link

arXiv 2606.10677 proposes Infini Memory, a text-based persistent memory architecture that structures an agent's long-term memory as topic documents — each topic acting as a semantic unit that collects related evidence and revises facts over time, rather than as an append-only log or opaque vector store. The framing is appealing for builders because the memory stays human-readable and editable, which matters for debugging and trust. It pairs naturally with the broader June push (HORMA, prior agent-memory characterization work) toward structured, navigable agent state.