Agent memory's new problem: remembering things that stopped being true
Yesterday's memory papers said plain retrieval beats fancy memory and that deleting a record doesn't really delete it. Four more from the same batch go after a quieter failure: memory that was right when written and is wrong when used.
Tether tests what happens when a fact in memory depends on something upstream that has since failed. An agent reading plain atomic-fact memory published 62 percent of those stale objects. Mem0 published 42 percent, A-Mem and Graphiti 9 percent. Tether records which versions a memory was derived from at write time and withdraws it at read time when the support breaks. It published 0.3 percent, with 37 percent fewer verification calls than checking everything. The weak spot: when related facts arrive in separate sessions, unsafe publication climbs back to 9.4 to 22.1 percent. https://arxiv.org/abs/2610.25123
AMV-Guard argues verification should depend on the action, not the memory. The same remembered address is fine for a routine shipment and should be re-confirmed before a valuable one. A value-of-information gate cut verification cost by 33.9 percent on ToolSandbox while keeping 100 percent safe-proceed recall, and by 21.3 percent on AppWorld. https://arxiv.org/abs/2610.31446
Two more round it out. A privacy study found persistent memory raises answer quality and leakage together, by a margin that widens over sessions. Disclosure feeds itself: once the agent mentions a personal detail, it becomes more likely to bring it up again, and scrubbing those mentions cut later disclosure about twice as much as single-turn effects predict. Telling the agent not to state or presuppose personal attributes brought leakage back to roughly no-memory levels while keeping most of the quality. https://arxiv.org/abs/2610.28315 And MemPABench, tracking interaction preferences over four simulated months, found Mem0, MemOS and simple RAG don't beat no memory at all. https://arxiv.org/abs/2610.30485
The pattern across the week: storing memories is solved. Knowing when a memory is still safe to act on is not, and that's where the real product is.
← Back to all articles
Tether tests what happens when a fact in memory depends on something upstream that has since failed. An agent reading plain atomic-fact memory published 62 percent of those stale objects. Mem0 published 42 percent, A-Mem and Graphiti 9 percent. Tether records which versions a memory was derived from at write time and withdraws it at read time when the support breaks. It published 0.3 percent, with 37 percent fewer verification calls than checking everything. The weak spot: when related facts arrive in separate sessions, unsafe publication climbs back to 9.4 to 22.1 percent. https://arxiv.org/abs/2610.25123
AMV-Guard argues verification should depend on the action, not the memory. The same remembered address is fine for a routine shipment and should be re-confirmed before a valuable one. A value-of-information gate cut verification cost by 33.9 percent on ToolSandbox while keeping 100 percent safe-proceed recall, and by 21.3 percent on AppWorld. https://arxiv.org/abs/2610.31446
Two more round it out. A privacy study found persistent memory raises answer quality and leakage together, by a margin that widens over sessions. Disclosure feeds itself: once the agent mentions a personal detail, it becomes more likely to bring it up again, and scrubbing those mentions cut later disclosure about twice as much as single-turn effects predict. Telling the agent not to state or presuppose personal attributes brought leakage back to roughly no-memory levels while keeping most of the quality. https://arxiv.org/abs/2610.28315 And MemPABench, tracking interaction preferences over four simulated months, found Mem0, MemOS and simple RAG don't beat no memory at all. https://arxiv.org/abs/2610.30485
The pattern across the week: storing memories is solved. Knowing when a memory is still safe to act on is not, and that's where the real product is.
Comments