Task configurations in factories, warehouses, and everyday physical environments change over time. Physical AI systems therefore need to incorporate new facts into scene memory while preserving valid information about equipment, objects, and operating constraints. We study updates to this parametric semantic-memory layer and present a joint evaluation of correction and preservation under finite update budgets. At fixed current fact count and memory parameter count, the protocol combines independent single-fact updates, budget trajectories, old-to-new answer diagnostics, and first stable success. Using a frozen 4B language model with late-layer LoRA, we evaluate seven preparation conditions and six update budgets on four-field synthetic semantic configurations. In two exposure-matched history conditions, increasing the budget from eight to 32 updates raises joint success from 6.25% to 94.79% and 91.15%, respectively, while exact obsolete-value outputs fall from 174/192 and 166/192 to zero. Constant32, which matches preparation steps but repeatedly trains on unchanged content during the first 32 events, achieves higher early joint success at eight steps (53.65%); at 32 steps, it reaches 85.42%, below both changing-history conditions. These results distinguish rapid correction under a small budget from simultaneous correctness of the new fact and the other three facts after further training. For both changing histories, most cases remain jointly correct at all subsequent measured points from step 16 onward. Output analysis shows that short-budget failures primarily retain the old value; repeated training on the new fact together with the other current facts progressively replaces these obsolete answers.
Figures & tables
Fig. 1: Exposure-matched order intervention. Each colored event changes one topic; the following gray event returns to the initial table. Both sequences contain the same 16 two-event loops and common tail. Colors denote topic roles, which rotate across configurations. The final state forks into independent probes, each changing one topic to v3. The schematic represents the two primary histories; reference branches are specified in the text.
Branch
Steps
Preparation
Interleaved
1,152
Alternating topic loops + tail
Blocked
1,152
Grouped topic loops + tail
Constant32
1,152
Constant table + tail
Short
128
Tail only
Refit32
32
Direct fit of current table
Refit128
128
Direct fit of current table
TABLE I: Endpoint preparation after the common 400-step initialization. Probe budgets are additional. The two primary histories match complete exposure; the other branches are descriptive references.
Branch
Current facts
Joint@8
New@8
Unchanged@8
Old@8
Joint@32
Facts@32
Interleaved
190/192
12/192
12/192
566/576
174/192
182/192
752/768
Blocked
180/192
12/192
14/192
546/576
166/192
175/192
740/768
Short
182/192
102/192
117/192
525/576
52/192
162/192
718/768
Constant32
185/192
103/192
117/192
545/576
60/192
164/192
722/768
Refit32
174/192
83/192
104/192
524/576
66/192
149/192
713/768
Refit128
180/192
89/192
105/192
535/576
70/192
151/192
715/768
TABLE II: Primary histories and descriptive references. Current facts: /192. Probe joint/new/old: /192; unchanged facts: /576; final facts: /768. References differ in exposure and/or preparation cost.
Fig. 2: Correction, preservation, and joint success at all predefined probe budgets. The vertical marker identifies the eight-step primary endpoint. The 0–100% axes provide a common scale. Constant32 maintains constant initial content with the same preparation update count as the primary histories. Lines connect measured checkpoints.
Fig. 3: Output and likelihood diagnostics. (A) Exact new, exact old, and all other changed-topic outputs; each row totals 192. (B) Median full-answer log-probability margins at measured checkpoints. Positive values favor the new answer; negative values favor the old answer. Constant32 starts with a more negative median and reaches a positive median earlier, showing a change in the ranking across budgets.
Fig. 4: Stable joint success on the fixed checkpoint grid. (A) Cumulative fraction reaching stable measured success by each budget. (B) First-success counts and cases right-censored at 32. Stability is defined as success at the first qualifying checkpoint and every later measured checkpoint through step 32.
Fig. 5: Paired history-order results. (A) All 16 configuration/initialization/edit-seed contrasts, each averaging four topics and three unseen wordings. (B) Overall and seed-group means. Dotted lines mark the prespecified effect-size screening thresholds. Opposite seed-level directions average to zero.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Fig. 6: Complete results. (A) Joint exact counts out of 192 at all predefined budgets, including every refit reference. (B) Current-endpoint exact mismatches partitioned by a string classifier into pronoun-only, music-label-only, and remaining lexical differences. The classification complements the original exact-match scores.
Memory-agent benchmarks often collapse retrieval and answer generation into one score, making it hard to tell whether a failure comes from missing an update or from not using an update that was already retrieved. We examine this distinction on the single-hop FactConsolidation task of MemoryAgentBench. With Qwen2.5-7B-Instruct on 6K-history BM25 accuracy drops from 45% to 32% when the retrieved evidence budget grows from 64 to 512 tokens, even though a gold-assisted post-hoc audit finds the newest statement in every retrieved set. Most of the errors occur on questions that also contain an obsolete statement version. Under a frozen prompt and equal four-statement evidence size, replacing one unrelated statement with the obsolete competing statement reduces accuracy from 100% to 44.6% for Qwen2.5 and from 100% to 87.8% for Qwen3-14B, with all 50 harmful paired flips classified as stale. A 32K-history replication likewise favors BM25-64 over BM25-512 (69% versus 58%). These results suggest that retrieving the correct fact is not enough. An obsolete competing statement can be more harmful than an unrelated piece of evidence of the same size.
Oleeviya Babu Poikarayil, Tony Basil Babu, Abdulrahman Nahhas +4
Faculty of Computer Science, Otto-von-Guericke-Universität Magdeburg, Germany · University of Kerala, India · Otto-von-Guericke-Universität Magdeburg, Germany +1
Learning from past experience benefits from two complementary forms of memory: episodic traces -- raw trajectories of what happened -- and consolidated abstractions distilled across many episodes into reusable, schema-like lessons. Recent agentic-memory systems pursue the consolidated form: an LLM rewrites past trajectories into a textual memory bank that it continuously updates with new interactions, promising self-improving agents without parameter updates. Yet we find that such consolidated memories produced by today's LLMs are often faulty even when derived from useful experiences. As consolidation proceeds, memory utility first rises, then degrades, and can fall below the no-memory baseline. More surprisingly, even when consolidating from ground-truth solutions, GPT-5.4 fails on 54% of a set of ARC-AGI problems it had previously solved without memory. We trace the regression to the consolidation step rather than the underlying experience: the same trajectories yield qualitatively different memories under different update schedules, and an episodic-only control that simply retains those trajectories remains competitive with the consolidators we test. In a controlled ARC-AGI Stream environment that exposes Retain, Delete, and Consolidate actions, agents preserve raw episodes by default and double the accuracy of their forced-consolidation counterparts; disabling consolidation entirely (episodic management only) matches this auto regime. Practically, robust agent memory should treat raw episodes as first-class evidence and gate consolidation explicitly rather than firing it after every interaction. Looking forward, reliable agentic memory will require LLMs that can consolidate without overwriting the evidence they depend on.
Dylan Zhang, Yanshan Lin, Zhengkun Wu +4
University of Illinois Urbana-Champaign · IIIS, Tsinghua University
Sequentially evolving LLM memory enables agents to reuse past experience, but existing systems usually deploy each locally generated memory update without checking whether it improves future behavior. As a result, updates that help the current task may overwrite useful knowledge, introduce over-specific rules, or bias the final memory toward recent examples. We propose Janus, a plug-in memory controller that decides whether to accept a candidate memory update or retain the previous memory. To make this decision efficient, Janus uses a Memory Momentum Trigger to identify suspicious deviations in the memory-update trajectory, and compares old and new memories on a compact hybrid evaluation set of coverage, boundary, and fresh tasks instead of replaying the full history. Janus is method-agnostic and wraps existing updaters without changing their update rules. Across six datasets, two backbone LLMs, and two memory updaters, Janus improves average accuracy by +2.7 to +4.6 points over the corresponding base updaters.
Zihan Chen, Songwei Dong, Chengshuai Shi +4
University of Virginia · Princeton University · University of Central Florida