Should your agent rewrite its own old memories when new ones arrive?

Wujiang Xu et al. · 1 authors · Rutgers University · 2025 · arXiv:2502.12110

Should your agent rewrite its own old memories when new ones arrive?

Opening

A business knowledge base that only grows eventually stops helping. Every week somebody pastes a new note, and every answer pulls back whatever happens to match the query, so yesterday’s phrasing competes with today’s and nobody ever goes back to fix the older entries. Most memory systems do exactly this: they store, they retrieve, and they leave the arrangement to chance. The question this paper asks is whether the agent should be allowed to reorganize its own store while it writes, the way a person rewrites their own notebook instead of only appending to it.

What the paper does

The paper proposes an agentic memory system built on the basic principles of the Zettelkasten method, turning memories into interconnected knowledge networks through dynamic indexing and linking. When a new memory arrives, the system writes a comprehensive note carrying structured attributes, then analyses historical memories to identify connections and establish links where similarities exist. Because new arrivals can also update the contextual representations of older notes, the store evolves rather than only grows. Empirical experiments run on six foundation models against existing state-of-the-art baselines.

Findings

  1. Storage without organization is the current default — existing memory systems can put memories away and fetch them back, but they do not arrange them, which is why the pile gets harder to use as it gets bigger.

Current memory systems enable basic storage and retrieval but lack sophisticated memory organization

  1. Each memory lands as a structured note — a new memory is not a raw string; it arrives with context, keywords, and tags that give later retrieval something to hold on to.

we generate a comprehensive note containing multiple structured attributes, including contextual descriptions, keywords, and tags

  1. New memories rewrite old ones — adding a memory can update the context and attributes of existing memories, so the network refines its understanding instead of only accumulating rows.

as new memories are integrated, they can trigger updates to the contextual representations and attributes of existing historical memories

  1. The gains hold across six models — the experiments report improvement over existing state-of-the-art baselines across six foundation models rather than on a single lucky model.

Empirical experiments on six foundation models show superior improvement against existing SOTA baselines

Caveat

Improvement is reported against existing memory baselines on six foundation models, with no task-level and no cost comparison, so the results do not establish that richer organization beats plain retrieval for real agent work.

What this means for 5ac

An SMB knowledge base changes every week, and this is the line between a usable memory and a pile of pasted chats: write every new fact as a note with context, keywords, and tags, and let a fresh note update the context of the older notes it extends or contradicts. The habit worth copying is the linking step, because links are what let a later query walk from today’s question to last month’s evidence. The caution to carry along with it is the caveat: the paper compares against memory baselines, not against a real task and a real invoice, so any store built on this pattern should be measured on task success and token cost before anyone calls it an improvement.

Source