Agent memory works now. The hard part is who it remembers for

Memory that respects the user boundary

MixUA, a preprint system for multi-user, multi-agent interaction, addresses the problem that makes shared agent memory dangerous: lessons learned from one user can leak to another. The system categorizes memory into granular scopes, user-agent pair, user, agent or global, and the authors report that it improves agent task pass rates by 16 points over no-memory baselines while limiting privacy leakage to 0.02% in multi-user environments. Multi-granularity memory scoping outperforms global memory systems by 5 points in task pass rates, the authors say, which means the privacy control is also a performance win, not a tax on it. m:…6046

The system distills trajectories into lessons scoped by user-agent pairs, users, agents or global access, and it is evaluated on CrossLearnBench, a new benchmark that tests agent performance on mixed-user tasks while probing for private fact leakage. The design insight is that memory scope is a first-class decision, not an afterthought. A lesson that is safe to share globally is different from a lesson that belongs to one user, and a system that cannot tell the difference will either leak private facts or refuse to learn anything at all. m:…6046

Reasoning memory beats action memory

Raindrop AI tested what kind of memory actually helps, using the tau-squared-bench telecom support benchmark across four open-weight models. Across all four models, the average share of tickets resolved increased from 33.5% without memory to 59.2% with action memory and 66.5% with reasoning memory. DeepSeek v4 Flash showed the largest gains: its resolution rate increased from 71% on tickets 2 to 8 to 93% in the final third of each sequence with reasoning memory. The company used Claude Opus 5.5 as the reviewer model to examine agent actions and outcomes. m:…2778

The catch is cost and scale. Access to reasoning can help further, Raindrop says, though the benefit varies by model and comes at a cost of approximately 30% more per ticket compared with action memory. Reasoning-aware reviewers can produce more specific, actionable lessons by identifying that an agent was guessing, which is exactly the kind of lesson that action-only memory cannot produce. Memory size eventually levels off as the reviewer model revises, consolidates and removes lessons, the company says, and the harness enforces a cap of 32 lessons and 14,400 bytes. Memory that grows without bound is a cost problem; memory that is curated is a product. m:…2778

Not every clever index helps

FlyHash, a fruit-fly-inspired hashing method for LLM agent memory, is a useful cautionary tale about optimizing the wrong metric. The authors report that FlyHash improves neighbor-structure preservation over SimHash under matched-bit protocols, with a mAP of 0.271 compared with 0.105 for SimHash at k=32, and that the quality advantage survives evaluation against genuine, non-circular agent-memory relevance labels. But empirical testing reveals that FlyHash does not provide practical memory or latency advantages over SimHash or HNSW at ordinary scales, and at equal bits FlyHash codes are larger and slower to compare than SimHash codes. m:…3138

HNSW dominates both FlyHash and SimHash at ordinary scale, the authors conclude, making FlyHash a quality-per-bit optimization rather than a drop-in replacement for existing indexing methods. The engineering lesson is to measure the metric that matters, latency and memory at the team's actual scale, rather than the metric that looks best in a paper. Oracle's Mem0 pattern makes a similar point from the storage side: Mem0 Python can use Oracle AI Database as a durable vector store for application memory, and the architecture separates memory operations from model calls to reduce prompt noise, with scoped retrieval and lifecycle management for memory records. m:…3138 m:…6880

Stale memory is the silent failure

Webz.io's study of news agents identifies the failure mode that the other papers mostly ignore: retirement. News agents often fail to retire stale conclusions because newer retrieval does not automatically invalidate older memory notes, and adding an accurate note to memory does not automatically remove the accessibility of an old, incorrect answer. A news agent needs a policy for retiring conclusions that applies to both stored briefings and memory notes, the author argues, because stored knowledge may become invalid as environments change. m:…4540

The proposed fix is claim versioning and dependency tracking: briefings should track dependencies on the notes used to generate them, new records should point to old records and provide a rationale for the replacement, and pending records should not be removed when status changes, since they remain valid for historical queries. Time-to-live is insufficient for determining claim validity, the author says, because announcements can be old but still correct. Oracle's pattern agrees on the mechanism: update and delete operations are necessary to keep memory aligned with current state rather than treating every record as permanent truth. Memory that cannot forget is a liability, not a feature. m:…4540 m:…6880