Memory Evolution on Retrieval
When a new memory links to existing neighbors, let those neighbors' own context, tags, and connections update in response, so the memory graph gets richer rather than merely bigger.
Where "agency" sits in a memory system is a design axis worth naming explicitly. Active retrieval and similar agentic-RAG systems put agency in the retrieval phase — the model autonomously decides when and what to fetch — while the underlying knowledge base itself stays static once written. Memory evolution puts agency somewhere else entirely: in storage, letting the memory structure itself change shape as new experience arrives, independent of whether or how it's later retrieved. A system can have either kind of agency without the other, or both; they answer different questions ("what do I fetch right now?" vs. "how should what I've already stored change in light of what just happened?").
Most memory systems are write-once: a stored note stays exactly as written until explicitly edited or evicted. Memory evolution breaks that assumption for the notes touched by embedding-plus-LLM linking: after a new memory's link set is computed, each neighbor note in that nearest-neighbor set is itself passed back through an LLM — given the new note, its other neighbors, and its own current content — to decide whether its context, keywords, or tags should update, and the evolved version replaces the original in the memory store.
The action space here is richer than a plain "update the summary": the evolution step can choose to strengthen a connection, merge two notes, prune a now-redundant one, and decide per neighbor whether to actually rewrite its content or merely reinforce the new link without touching the text — not just "leave unchanged" versus "overwrite."
The effect compounds over time: every new interaction can revise several existing notes' framing, not just add one more note to the pile, so the network's organization becomes progressively richer as new experience interacts with old notes — closer to how encountering a new example can sharpen how you remember several older, related examples, rather than simply appending the new example to a list. Contrast this with episodic reflection memory's fixed-size buffer of self-contained lesson entries, which never rewrites earlier entries in light of a later one — memory evolution is specifically the mechanism that lets earlier notes change, not just the newest one get appended.
Sources
- A-Mem (Xu et al.), §3.3