542 lines
28 KiB
JSON
542 lines
28 KiB
JSON
{
|
||
"built": "2026-10-25",
|
||
"query": "all:\"agentic memory\"",
|
||
"feed_updated": "2026-10-25T00:40:42Z",
|
||
"feed_total_results": 583,
|
||
"papers": [
|
||
{
|
||
"arxiv_id": "2610.29671",
|
||
"title": "Uni-Mem: Learning a Long-Horizon Agentic Memory Manager with Unified Online Policy Optimization",
|
||
"bytes": 90083,
|
||
"chars": 89293,
|
||
"sha256": "b83a772fb6a1340b2440359bec2bf3752d583d86e213806c5dbf7a648d947bdd",
|
||
"claims": 9
|
||
},
|
||
{
|
||
"arxiv_id": "2610.29823",
|
||
"title": "MemoryTax: The Bitter Lesson of Agent Memory",
|
||
"bytes": 23897,
|
||
"chars": 23627,
|
||
"sha256": "ed08ec2380a6c1479c1cbc0534cc2cf154f92aa057216398055f034f91abd9b5",
|
||
"claims": 9
|
||
},
|
||
{
|
||
"arxiv_id": "2610.29869",
|
||
"title": "Compaction Laundering: Provenance Destruction in Context-Compacting Coding Agents",
|
||
"bytes": 92858,
|
||
"chars": 91936,
|
||
"sha256": "e6f78f508c12d7a304f7fd36eb5f05d60095a426b189ab4a5cd817c341791b54",
|
||
"claims": 7
|
||
},
|
||
{
|
||
"arxiv_id": "2610.30485",
|
||
"title": "Why Bother Me Again? Benchmarking Personal Agent Memory for Interaction Preferences",
|
||
"bytes": 350788,
|
||
"chars": 340805,
|
||
"sha256": "48e8b2b9c7defc653b8c7577e9e6029a5f9873cbaabe0f6b54f7ca84b08a1906",
|
||
"claims": 9
|
||
},
|
||
{
|
||
"arxiv_id": "2610.30755",
|
||
"title": "Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought",
|
||
"bytes": 195193,
|
||
"chars": 193677,
|
||
"sha256": "2246678ceaf7a42ca8104b5067a9d548b0795391af40142b62e3f1638288daac",
|
||
"claims": 10
|
||
}
|
||
],
|
||
"themes": {
|
||
"T1": "Is designed memory needed at all?",
|
||
"T2": "Cost, ceiling, and scale",
|
||
"T3": "Provenance: what compaction destroys and what writers omit",
|
||
"T4": "What raw evidence must be preserved",
|
||
"T5": "Interaction preferences are a different problem"
|
||
},
|
||
"claims": [
|
||
{
|
||
"id": "R1",
|
||
"paper": "2610.30755",
|
||
"theme": "T1",
|
||
"claim": "Raw-RAG ranks first among compared memory systems in six of seven benchmark-model settings.",
|
||
"value": "six of seven",
|
||
"note": "The headline result: the naive baseline wins most settings.",
|
||
"anchor": "Raw-RAG ranks first among the compared memory systems in six",
|
||
"quote": "Raw-RAG ranks first among the compared memory systems in six of seven benchmark–model settings, outperforming the strongest compared agentic memory system by 0.13–5.53 points. We investigate why raw-history ret",
|
||
"corpus_sha256": "2246678ceaf7a42ca8104b5067a9d548b0795391af40142b62e3f1638288daac"
|
||
},
|
||
{
|
||
"id": "R2",
|
||
"paper": "2610.30755",
|
||
"theme": "T1",
|
||
"claim": "Raw-RAG beats the strongest compared agentic memory system by 0.13–5.53 points.",
|
||
"value": "0.13–5.53 points",
|
||
"note": "Margin is small on LoCoMo/LongMemEval, large on MemoryAgentBench.",
|
||
"anchor": "outperforming the strongest compared agentic memory system by 0.13",
|
||
"quote": "outperforming the strongest compared agentic memory system by 0.13–5.53 points. We investigate why raw-history retrieval remains c",
|
||
"corpus_sha256": "2246678ceaf7a42ca8104b5067a9d548b0795391af40142b62e3f1638288daac"
|
||
},
|
||
{
|
||
"id": "R3",
|
||
"paper": "2610.30755",
|
||
"theme": "T4",
|
||
"claim": "Mem0's UPDATE+DELETE account for only 4.39% of 33,902 LoCoMo write decisions.",
|
||
"value": "4.39%",
|
||
"note": "Mutation is rare in practice -- most proposed writes are duplicates or malformed.",
|
||
"anchor": "together account for only 4.39% of the 33,902 write decisions on LoCoMo",
|
||
"quote": "together account for only 4.39% of the 33,902 write decisions on LoCoMo\nand 2.46% of the 428,858 write decisions on LongMemEval\n(Figure 2); ",
|
||
"corpus_sha256": "2246678ceaf7a42ca8104b5067a9d548b0795391af40142b62e3f1638288daac"
|
||
},
|
||
{
|
||
"id": "R4",
|
||
"paper": "2610.30755",
|
||
"theme": "T4",
|
||
"claim": "Replacing UPDATE/DELETE with append-only ADD/NONE raises LoCoMo accuracy 67.36% -> 69.42%.",
|
||
"value": "67.36% to 69.42%",
|
||
"note": "Explicit mutation gives no accuracy gain; the simpler store is better.",
|
||
"anchor": "Replacing UPDATE with append-only ADD and DELETE with NONE raises",
|
||
"quote": "Replacing UPDATE with append-only ADD and DELETE with NONE raises\nLoCoMo accuracy from 67.36% to 69.42%, without reducing any category score\n(Table 2). Explicit mutation ",
|
||
"corpus_sha256": "2246678ceaf7a42ca8104b5067a9d548b0795391af40142b62e3f1638288daac"
|
||
},
|
||
{
|
||
"id": "R5",
|
||
"paper": "2610.30755",
|
||
"theme": "T4",
|
||
"claim": "Removing reranking lowers MemoryAgentBench Total by 3.92 points.",
|
||
"value": "3.92 points",
|
||
"note": "Reranking is the one component worth keeping.",
|
||
"anchor": "Removing reranking lowers Total by 3.92 points",
|
||
"quote": "Removing reranking lowers Total by 3.92 points (pHolm=0.0012p_{\\mathrm{Holm}}=0.0012),\nwhile returning only one item lowers it by 11.73 points. Final-kk\nvalues of 5, 10, 30, and 60 stay with",
|
||
"corpus_sha256": "2246678ceaf7a42ca8104b5067a9d548b0795391af40142b62e3f1638288daac"
|
||
},
|
||
{
|
||
"id": "R6",
|
||
"paper": "2610.30755",
|
||
"theme": "T4",
|
||
"claim": "Returning only one item lowers Total by 11.73 points.",
|
||
"value": "11.73 points",
|
||
"note": "Retrieval depth matters far more than chunk size.",
|
||
"anchor": "while returning only one item lowers it by 11.73 points",
|
||
"quote": "while returning only one item lowers it by 11.73 points. Final-kk\nvalues of 5, 10, 30, and 60 stay within 1.3 points of the baseli",
|
||
"corpus_sha256": "2246678ceaf7a42ca8104b5067a9d548b0795391af40142b62e3f1638288daac"
|
||
},
|
||
{
|
||
"id": "R7",
|
||
"paper": "2610.30755",
|
||
"theme": "T4",
|
||
"claim": "Top-120 retrieval loses 14.57 points.",
|
||
"value": "14.57 points",
|
||
"note": "Driven by long-range understanding and conflict resolution.",
|
||
"anchor": "Top-120 loses 14.57 points",
|
||
"quote": "Top-120 loses 14.57 points, driven by\nlong-range understanding and conflict resolution. Chunk sizes from 128 to 1,024\ndo not signi",
|
||
"corpus_sha256": "2246678ceaf7a42ca8104b5067a9d548b0795391af40142b62e3f1638288daac"
|
||
},
|
||
{
|
||
"id": "R8",
|
||
"paper": "2610.30755",
|
||
"theme": "T1",
|
||
"claim": "No memory-management variant improves on Raw-RAG (48.32%).",
|
||
"value": "48.32%",
|
||
"note": "RQ2 answer: added management does not help.",
|
||
"anchor": "None improves Raw-RAG (48.32%)",
|
||
"quote": "None improves Raw-RAG (48.32%) on the\nsame 14-source MemoryAgentBench Total (Table 6, orange rows).\nThe exact construction and retrieval procedures ar",
|
||
"corpus_sha256": "2246678ceaf7a42ca8104b5067a9d548b0795391af40142b62e3f1638288daac"
|
||
},
|
||
{
|
||
"id": "R9",
|
||
"paper": "2610.30755",
|
||
"theme": "T1",
|
||
"claim": "Atomic-fact extraction collapses test-time learning from 47.68% to 15.37%.",
|
||
"value": "47.68% to 15.37%",
|
||
"note": "The most damaging of the three management operations.",
|
||
"anchor": "Atomic-fact extraction reduces test-time learning from 47.68% to 15.37%",
|
||
"quote": "Atomic-fact extraction reduces test-time learning from 47.68% to 15.37%\nand Total by 19.62 points, consistent with loss or disrupt",
|
||
"corpus_sha256": "2246678ceaf7a42ca8104b5067a9d548b0795391af40142b62e3f1638288daac"
|
||
},
|
||
{
|
||
"id": "R10",
|
||
"paper": "2610.30755",
|
||
"theme": "T2",
|
||
"claim": "Full-context beats Raw-RAG by 19.98 points on DeepSeek-V4-Flash MEME.",
|
||
"value": "19.98 points",
|
||
"note": "The one setting where the naive baseline is not first.",
|
||
"anchor": "Full-context leads Raw-RAG by 19.98 points",
|
||
"quote": "Full-context leads Raw-RAG by 19.98 points in that setting. Raw-RAG\nis competitive across these tasks despite its simpler ingestio",
|
||
"corpus_sha256": "2246678ceaf7a42ca8104b5067a9d548b0795391af40142b62e3f1638288daac"
|
||
},
|
||
{
|
||
"id": "P1",
|
||
"paper": "2610.30485",
|
||
"theme": "T5",
|
||
"claim": "MemPABench has four simulated users holding 14 preferences.",
|
||
"value": "14 preferences",
|
||
"note": "Preferences span work and personal contexts.",
|
||
"anchor": "Four grounded simulated users hold 14 preferences",
|
||
"quote": "Four grounded simulated users hold 14 preferences across work and personal contexts. Their feedback spans 99 sessions over four months, with",
|
||
"corpus_sha256": "48e8b2b9c7defc653b8c7577e9e6029a5f9873cbaabe0f6b54f7ca84b08a1906"
|
||
},
|
||
{
|
||
"id": "P2",
|
||
"paper": "2610.30485",
|
||
"theme": "T5",
|
||
"claim": "Feedback spans 99 sessions over four months.",
|
||
"value": "99 sessions over four months",
|
||
"note": "Five event-driven preference changes per user.",
|
||
"anchor": "Their feedback spans 99 sessions over four months",
|
||
"quote": "Their feedback spans 99 sessions over four months, with five event-driven changes per user testing memory updates. We evaluate pre",
|
||
"corpus_sha256": "48e8b2b9c7defc653b8c7577e9e6029a5f9873cbaabe0f6b54f7ca84b08a1906"
|
||
},
|
||
{
|
||
"id": "P3",
|
||
"paper": "2610.30485",
|
||
"theme": "T5",
|
||
"claim": "No Memory reaches 49.6% PSA against 45.2% random.",
|
||
"value": "49.6%",
|
||
"note": "Defaults alone nearly match chance-plus-margin.",
|
||
"anchor": "No Memory, where the PA selects by its defaults, reaches 49.6%",
|
||
"quote": "No Memory, where the PA selects by its defaults, reaches 49.6%, compared with 45.2% for random selection (Appendix A.1). Only File memory adds a clear gain. It raises PSA to 59.2%, a mean ga",
|
||
"corpus_sha256": "48e8b2b9c7defc653b8c7577e9e6029a5f9873cbaabe0f6b54f7ca84b08a1906"
|
||
},
|
||
{
|
||
"id": "P4",
|
||
"paper": "2610.30485",
|
||
"theme": "T5",
|
||
"claim": "File memory raises PSA to 59.2%, +9.6 points over No Memory.",
|
||
"value": "59.2%",
|
||
"note": "The only condition with a clear gain; CI [+3.9, +15.6].",
|
||
"anchor": "It raises PSA to 59.2%",
|
||
"quote": "It raises PSA to 59.2%, a mean gain of 9.6 points over No Memory (95% cluster-bootstrap interval [+3.9, +15.6]). PERMA (Liu et al., 2026), a benchmark of content preferen",
|
||
"corpus_sha256": "48e8b2b9c7defc653b8c7577e9e6029a5f9873cbaabe0f6b54f7ca84b08a1906"
|
||
},
|
||
{
|
||
"id": "P5",
|
||
"paper": "2610.30485",
|
||
"theme": "T5",
|
||
"claim": "Mem0, MemOS and Simple RAG stay at No Memory level (PSA 50.0/50.0/49.9).",
|
||
"value": "50.0%, 50.0%, and 49.9%",
|
||
"note": "Retrieval-based memory does not beat having none.",
|
||
"anchor": "with PSA of 50.0%, 50.0%, and 49.9%",
|
||
"quote": "with PSA of 50.0%, 50.0%, and 49.9%. The intervals of all conditions and their comparisons with No Memory on each user–backbone pa",
|
||
"corpus_sha256": "48e8b2b9c7defc653b8c7577e9e6029a5f9873cbaabe0f6b54f7ca84b08a1906"
|
||
},
|
||
{
|
||
"id": "P6",
|
||
"paper": "2610.30485",
|
||
"theme": "T5",
|
||
"claim": "File memory PET is 25.4%, more than twice No Memory's 11.2%.",
|
||
"value": "25.4%",
|
||
"note": "But its CS does not improve (14.4% vs 21.8%): context is not tracked.",
|
||
"anchor": "Its PET is 25.4%",
|
||
"quote": "Its PET is 25.4%, more than twice the 11.2% of No Memory, and it is higher on every backbone. However, the CS does not improve, at 14.4% against 21.8% (Appendix G.3). In ",
|
||
"corpus_sha256": "48e8b2b9c7defc653b8c7577e9e6029a5f9873cbaabe0f6b54f7ca84b08a1906"
|
||
},
|
||
{
|
||
"id": "P7",
|
||
"paper": "2610.30485",
|
||
"theme": "T5",
|
||
"claim": "Oracle lifts correct selection 33%->96% but BRS>=4 only 24%->59%.",
|
||
"value": "33% to 96%",
|
||
"note": "Knowing the preference does not override the model's interaction habits.",
|
||
"anchor": "Oracle raises correct selections of the five settings",
|
||
"quote": "Oracle raises correct selections of the five settings from 33% to 96%, but the share of replies with a BRS of 4 or higher rises only from 24% to 59%. Even with the IX Wrapper, the PA still writes too much. Afte",
|
||
"corpus_sha256": "48e8b2b9c7defc653b8c7577e9e6029a5f9873cbaabe0f6b54f7ca84b08a1906"
|
||
},
|
||
{
|
||
"id": "P8",
|
||
"paper": "2610.30485",
|
||
"theme": "T5",
|
||
"claim": "Information Flow is the hardest dimension, PSA 34–44%.",
|
||
"value": "34–44%",
|
||
"note": "Expression Style (57-69%) and Disclosure (55-58%) are easier.",
|
||
"anchor": "Information Flow is the hardest under every condition",
|
||
"quote": "Information Flow is the hardest under every condition (Figure 8). Its PSA stays at 34–44%, the lowest of the four, while Expression Style (57–69%) and Disclosur",
|
||
"corpus_sha256": "48e8b2b9c7defc653b8c7577e9e6029a5f9873cbaabe0f6b54f7ca84b08a1906"
|
||
},
|
||
{
|
||
"id": "P9",
|
||
"paper": "2610.30485",
|
||
"theme": "T5",
|
||
"claim": "Anonymizing users drops shortcut accuracy 53.2% -> 46.4%, near the 45.2% chance.",
|
||
"value": "53.2%",
|
||
"note": "Confirms the benchmark measures memory, not name-guessing.",
|
||
"anchor": "reaches 53.2% accuracy from the source identity but 46.4%",
|
||
"quote": "reaches 53.2% accuracy from the source identity but 46.4% from the anonymized profile, close to the 45.2% chance level (Figure 10; Appendix ",
|
||
"corpus_sha256": "48e8b2b9c7defc653b8c7577e9e6029a5f9873cbaabe0f6b54f7ca84b08a1906"
|
||
},
|
||
{
|
||
"id": "M1",
|
||
"paper": "2610.29823",
|
||
"theme": "T2",
|
||
"claim": "Study scope: one LoCoMo conversation, 419 turns, 152 questions.",
|
||
"value": "419 turns, 152 questions",
|
||
"note": "Deliberately narrow: one conversation, one run per cell.",
|
||
"anchor": "On one LoCoMo conversation (419 turns, 152 questions)",
|
||
"quote": "On one LoCoMo conversation (419 turns, 152 questions), we run four memory methods (full context, BM25, Mem0, Zep) and three coding agents (Pi, Codex, Claude Cod",
|
||
"corpus_sha256": "ed08ec2380a6c1479c1cbc0534cc2cf154f92aa057216398055f034f91abd9b5"
|
||
},
|
||
{
|
||
"id": "M2",
|
||
"paper": "2610.29823",
|
||
"theme": "T2",
|
||
"claim": "With GPT-4o-mini, Mem0 scores 78% vs 74% for full context and BM25.",
|
||
"value": "78%",
|
||
"note": "Agent memory is a patch for a weak model.",
|
||
"anchor": "With GPT-4o-mini, Mem0 scores highest",
|
||
"quote": "With GPT-4o-mini, Mem0 scores highest. It scores 78%, against 74% for both full context and BM25; Zep scores 43%. The agents score 63% (Pi), 28% (Codex) and 29%",
|
||
"corpus_sha256": "ed08ec2380a6c1479c1cbc0534cc2cf154f92aa057216398055f034f91abd9b5"
|
||
},
|
||
{
|
||
"id": "M3",
|
||
"paper": "2610.29823",
|
||
"theme": "T2",
|
||
"claim": "With GPT-6 Luna, full context scores 92%, BM25 82%, Mem0 80%, Zep 51%.",
|
||
"value": "92%",
|
||
"note": "The patch stops helping as the model gets stronger.",
|
||
"anchor": "With GPT-6 Luna, full context scores 92%",
|
||
"quote": "With GPT-6 Luna, full context scores 92%, BM25 82%, Mem0 80% and Zep 51%; Zep’s graph holds 128 facts from the 19 sessions, against 195 with Claude Opus 5.5. With Claude Opus 5.5, Mem0 and full contex",
|
||
"corpus_sha256": "ed08ec2380a6c1479c1cbc0534cc2cf154f92aa057216398055f034f91abd9b5"
|
||
},
|
||
{
|
||
"id": "M4",
|
||
"paper": "2610.29823",
|
||
"theme": "T2",
|
||
"claim": "With Claude Opus 5.5, agents score highest: Pi 99%, Codex 98%, Claude Code 97%.",
|
||
"value": "Pi 99%, Codex 98% and Claude Code 97%",
|
||
"note": "Searching the raw record beats reading all of it at the top end.",
|
||
"anchor": "With Claude Opus 5.5, Mem0 and full context both score 95%",
|
||
"quote": "With Claude Opus 5.5, Mem0 and full context both score 95% and Zep 87%, and the agents score highest: Pi 99%, Codex 98% and Claude Code 97%, a lead that comes mostly from",
|
||
"corpus_sha256": "ed08ec2380a6c1479c1cbc0534cc2cf154f92aa057216398055f034f91abd9b5"
|
||
},
|
||
{
|
||
"id": "M5",
|
||
"paper": "2610.29823",
|
||
"theme": "T2",
|
||
"claim": "Pi scores 99% for $6.95/rollout vs full context 95% for $17.84.",
|
||
"value": "$6.95",
|
||
"note": "Less than half the cost for a higher score.",
|
||
"anchor": "Pi scores 99% for $6.95 a rollout",
|
||
"quote": "Pi scores 99% for $6.95 a rollout, while full context with the same model scores 95% for $17.84. Pi reads about 10,000 input tokens a question, 7,900 of them ca",
|
||
"corpus_sha256": "ed08ec2380a6c1479c1cbc0534cc2cf154f92aa057216398055f034f91abd9b5"
|
||
},
|
||
{
|
||
"id": "M6",
|
||
"paper": "2610.29823",
|
||
"theme": "T4",
|
||
"claim": "On the 88 questions without photo evidence, Pi scores 100%.",
|
||
"value": "88 questions without photo evidence",
|
||
"note": "Pi matches full context once captions are removed from the comparison.",
|
||
"anchor": "On the 88 questions without photo evidence",
|
||
"quote": "On the 88 questions without photo evidence, Pi scores 100%, and full context, Mem0, Codex and Claude Code 99% each (Table 4). With the same input, then, Pi matc",
|
||
"corpus_sha256": "ed08ec2380a6c1479c1cbc0534cc2cf154f92aa057216398055f034f91abd9b5"
|
||
},
|
||
{
|
||
"id": "M7",
|
||
"paper": "2610.29823",
|
||
"theme": "T2",
|
||
"claim": "Pi reads ~10,000 input tokens/question (7,900 cached) vs 28,900 for full context.",
|
||
"value": "10,000 input tokens",
|
||
"note": "The efficiency mechanism behind the cost lead.",
|
||
"anchor": "Pi reads about 10,000 input tokens a question",
|
||
"quote": "Pi reads about 10,000 input tokens a question, 7,900 of them cached, against 28,900 for full context. Codex and Claude Code score 98% and 97% for $25.",
|
||
"corpus_sha256": "ed08ec2380a6c1479c1cbc0534cc2cf154f92aa057216398055f034f91abd9b5"
|
||
},
|
||
{
|
||
"id": "M8",
|
||
"paper": "2610.29823",
|
||
"theme": "T3",
|
||
"claim": "Writer ceiling: on Zep's GPT-4o-mini graph readers score 43/43/45%; on Opus 5.5's, 66/71/87%.",
|
||
"value": "43%, 43% and 45%",
|
||
"note": "A stronger reader cannot recover what a weaker writer left out.",
|
||
"anchor": "GPT-4o-mini graph, GPT-4o-mini, GPT-6 Luna and Claude Opus 5.5 score 43%, 43% and 45%",
|
||
"quote": "GPT-4o-mini graph, GPT-4o-mini, GPT-6 Luna and Claude Opus 5.5 score 43%, 43% and 45%; on the graph Claude Opus 5.5 wrote, 66%, 71% and 87%. Mem0 keeps the order with smaller gaps: 78%, 83% and 84% from GPT-4o-",
|
||
"corpus_sha256": "ed08ec2380a6c1479c1cbc0534cc2cf154f92aa057216398055f034f91abd9b5"
|
||
},
|
||
{
|
||
"id": "M9",
|
||
"paper": "2610.29823",
|
||
"theme": "T3",
|
||
"claim": "Pi's one miss is a LoCoMo answer-key error, not a memory failure.",
|
||
"value": "151 questions",
|
||
"note": "The model is right and the benchmark key is wrong.",
|
||
"anchor": "Its one miss is LoCoMo",
|
||
"quote": "Its one miss is LoCoMo’s mistake. The question asks what kind of painting Caroline shared with Melanie on 13 October 2023, and the answer key says an abstract painting with blue streaks. In the conversation it is Melanie who shares that painting; Caroline shares a drawing. Pi says so and is marked wrong: “Caroline shared a drawing, not a painting. … The paintings shared that day (a sunset and an abstract piece) were Melanie’s.” With Claude Opus 5.5, full context, Codex and Claude Code point out the same mistake and are marked wrong too. On the other 151 questions, Pi with Claude Opus",
|
||
"corpus_sha256": "ed08ec2380a6c1479c1cbc0534cc2cf154f92aa057216398055f034f91abd9b5"
|
||
},
|
||
{
|
||
"id": "C1",
|
||
"paper": "2610.29869",
|
||
"theme": "T3",
|
||
"claim": "Forged approval launders into memory 87–100% of the time across five flattening harnesses.",
|
||
"value": "87–100%",
|
||
"note": "Up to a ~908,000-token context window.",
|
||
"anchor": "the forged approval launders into the",
|
||
"quote": "the forged approval launders into the\nagent’s memory 87–100% of the time, up to a\n~908,000-token context window; across four role-typed\nharnesses (Codex, Gemini CLI, Qwen Code, Continue) the identical attack\nsu",
|
||
"corpus_sha256": "e6f78f508c12d7a304f7fd36eb5f05d60095a426b189ab4a5cd817c341791b54"
|
||
},
|
||
{
|
||
"id": "C2",
|
||
"paper": "2610.29869",
|
||
"theme": "T3",
|
||
"claim": "The identical attack succeeds 0% of the time across four role-typed harnesses.",
|
||
"value": "0% of the time",
|
||
"note": "Role-typing is a built-in control.",
|
||
"anchor": "succeeds 0% of the time",
|
||
"quote": "succeeds 0% of the time. We follow the attack to executed consequences.\nA laundered standing approval drives real tool-running age",
|
||
"corpus_sha256": "e6f78f508c12d7a304f7fd36eb5f05d60095a426b189ab4a5cd817c341791b54"
|
||
},
|
||
{
|
||
"id": "C3",
|
||
"paper": "2610.29869",
|
||
"theme": "T3",
|
||
"claim": "Live in the real pi TUI the attack runs at ~50% against a 0% paired clean arm.",
|
||
"value": "~50%",
|
||
"note": "Runtime evidence, not a simulation.",
|
||
"anchor": "live in the real pi TUI at ~50%",
|
||
"quote": "live in the real pi TUI at ~50%\nagainst a 0% paired clean arm. Straiker [1] described this\nconsequence without runtime evidence, a",
|
||
"corpus_sha256": "e6f78f508c12d7a304f7fd36eb5f05d60095a426b189ab4a5cd817c341791b54"
|
||
},
|
||
{
|
||
"id": "C4",
|
||
"paper": "2610.29869",
|
||
"theme": "T3",
|
||
"claim": "A code-level provenance check catches 77% of laundering, demoting 2% of legitimate content.",
|
||
"value": "77%",
|
||
"note": "The only defense that works; prompt-level defenses cannot restore provenance.",
|
||
"anchor": "we measure a code-level provenance check that catches 77% of",
|
||
"quote": "we measure a code-level provenance check that catches 77% of\nlaundering while demoting 2% of legitimate summary content. The defect\nalso has a straightforward f",
|
||
"corpus_sha256": "e6f78f508c12d7a304f7fd36eb5f05d60095a426b189ab4a5cd817c341791b54"
|
||
},
|
||
{
|
||
"id": "C5",
|
||
"paper": "2610.29869",
|
||
"theme": "T3",
|
||
"claim": "In-prompt anti-injection rules stop 13–25%; published classifier defenses report 67% at best.",
|
||
"value": "13–25%",
|
||
"note": "Baselines are far weaker than the code-level check.",
|
||
"anchor": "stops 13",
|
||
"quote": "stops 13–25% on the same data, and published classifier\ndefenses report 67% at best. The defect also has a straightforward\nfix on the flatte",
|
||
"corpus_sha256": "e6f78f508c12d7a304f7fd36eb5f05d60095a426b189ab4a5cd817c341791b54"
|
||
},
|
||
{
|
||
"id": "C6",
|
||
"paper": "2610.29869",
|
||
"theme": "T3",
|
||
"claim": "Laundering still succeeds 100% of the time at every context size.",
|
||
"value": "100%",
|
||
"note": "Size does not attenuate the attack.",
|
||
"anchor": "The laundering still succeeds 100% of the",
|
||
"quote": "The laundering still succeeds 100% of the\ntime at every size, and the executed rate, near-100% on the short\ntranscripts, settles to ~20–40% in a full window. Th",
|
||
"corpus_sha256": "e6f78f508c12d7a304f7fd36eb5f05d60095a426b189ab4a5cd817c341791b54"
|
||
},
|
||
{
|
||
"id": "C7",
|
||
"paper": "2610.29869",
|
||
"theme": "T3",
|
||
"claim": "Executed rate settles to ~20–40% in a full window.",
|
||
"value": "~20–40%",
|
||
"note": "Real agents compact at this size, so this is the realistic case.",
|
||
"anchor": "settles to ~20",
|
||
"quote": "settles to ~20–40% in a full window. The\nshort runs therefore measure the attack at its most favorable, an\nupper bound on the exec",
|
||
"corpus_sha256": "e6f78f508c12d7a304f7fd36eb5f05d60095a426b189ab4a5cd817c341791b54"
|
||
},
|
||
{
|
||
"id": "U1",
|
||
"paper": "2610.29671",
|
||
"theme": "T1",
|
||
"claim": "Uni-Mem reaches an overall LoCoMo JJ of 88.80%.",
|
||
"value": "88.80%",
|
||
"note": "Best across all four question types.",
|
||
"anchor": "Uni-Mem achieves the best performance across all four question types",
|
||
"quote": "Uni-Mem achieves the best performance across all four question types, with an overall JJ score of 88.80%.\nIt surpasses the strongest baseline, Memory-R2 (OSS), by 10.93 percentage points and Mem-T by 12.31 points, with g",
|
||
"corpus_sha256": "b83a772fb6a1340b2440359bec2bf3752d583d86e213806c5dbf7a648d947bdd"
|
||
},
|
||
{
|
||
"id": "U2",
|
||
"paper": "2610.29671",
|
||
"theme": "T1",
|
||
"claim": "Uni-Mem surpasses Memory-R2 (OSS) by 10.93 points and Mem-T by 12.31.",
|
||
"value": "10.93 percentage points",
|
||
"note": "Gains of 12.93 and 9.29 on multi-hop and temporal.",
|
||
"anchor": "It surpasses the strongest baseline, Memory-R2 (OSS), by 10.93",
|
||
"quote": "It surpasses the strongest baseline, Memory-R2 (OSS), by 10.93 percentage points and Mem-T by 12.31 points, with gains of 12.93 and 9.29 points over the strongest competing results on multi-hop and temporal que",
|
||
"corpus_sha256": "b83a772fb6a1340b2440359bec2bf3752d583d86e213806c5dbf7a648d947bdd"
|
||
},
|
||
{
|
||
"id": "U3",
|
||
"paper": "2610.29671",
|
||
"theme": "T1",
|
||
"claim": "Uni-Mem still reaches 80.99% without SFT, +3.12% over the strongest baseline.",
|
||
"value": "80.99%",
|
||
"note": "RL alone is enough to beat the baselines.",
|
||
"anchor": "Uni-Mem still achieves 80.99% without SFT",
|
||
"quote": "Uni-Mem still achieves 80.99% without SFT, outperforming the strongest baseline by 3.12%.\nThese results support the effectiveness of jointly",
|
||
"corpus_sha256": "b83a772fb6a1340b2440359bec2bf3752d583d86e213806c5dbf7a648d947bdd"
|
||
},
|
||
{
|
||
"id": "U4",
|
||
"paper": "2610.29671",
|
||
"theme": "T1",
|
||
"claim": "ECA improves overall JJ by 1.29 and 3.03 points without and with SFT.",
|
||
"value": "1.29 and 3.03 percentage points",
|
||
"note": "Event-wise credit assignment localizes sparse feedback.",
|
||
"anchor": "improves overall JJ by 1.29 and 3.03 percentage points",
|
||
"quote": "improves overall JJ by 1.29 and 3.03 percentage points without and with SFT, respectively. These gains support the effectiveness o",
|
||
"corpus_sha256": "b83a772fb6a1340b2440359bec2bf3752d583d86e213806c5dbf7a648d947bdd"
|
||
},
|
||
{
|
||
"id": "U5",
|
||
"paper": "2610.29671",
|
||
"theme": "T1",
|
||
"claim": "Stage-Decoupled GRPO improves overall JJ by 2.48 points without SFT.",
|
||
"value": "2.48 percentage points",
|
||
"note": "Temporal accuracy rises 10.55 points in this setting.",
|
||
"anchor": "improves overall JJ by 2.48 percentage points without SFT",
|
||
"quote": "improves overall JJ by 2.48 percentage points without SFT. In particular, temporal accuracy rises by 10.55 points in this setting, and multi",
|
||
"corpus_sha256": "b83a772fb6a1340b2440359bec2bf3752d583d86e213806c5dbf7a648d947bdd"
|
||
},
|
||
{
|
||
"id": "U6",
|
||
"paper": "2610.29671",
|
||
"theme": "T1",
|
||
"claim": "Trained construction alone improves overall JJ by 8.18 points.",
|
||
"value": "8.18 percentage points",
|
||
"note": "Each capability contributes even paired with its untrained counterpart.",
|
||
"anchor": "improves the overall JJ score by 8.18 percentage points",
|
||
"quote": "improves the overall JJ score by 8.18 percentage points, demonstrating the increased downstream utility of the constructed memorie",
|
||
"corpus_sha256": "b83a772fb6a1340b2440359bec2bf3752d583d86e213806c5dbf7a648d947bdd"
|
||
},
|
||
{
|
||
"id": "U7",
|
||
"paper": "2610.29671",
|
||
"theme": "T1",
|
||
"claim": "Extending update blocks 1:1 -> 3:2 reduces overall JJ from 88.80% to 83.75%.",
|
||
"value": "83.75%",
|
||
"note": "Frequent alternation matters as much as balanced allocation.",
|
||
"anchor": "the overall JJ score from 88.80",
|
||
"quote": "the overall JJ score from 88.80%88.80\\% to 83.75%83.75\\%,\nindicating that less frequent alternation hinders\nshared-polic",
|
||
"corpus_sha256": "b83a772fb6a1340b2440359bec2bf3752d583d86e213806c5dbf7a648d947bdd"
|
||
},
|
||
{
|
||
"id": "U8",
|
||
"paper": "2610.29671",
|
||
"theme": "T2",
|
||
"claim": "SFT followed by RL improves the 4B manager's JJ from 76.58% to 88.80%.",
|
||
"value": "76.58% to 88.80%",
|
||
"note": "A small trained manager beats Qwen3.5-27B (87.05%) and 397B-A17B (85.86%).",
|
||
"anchor": "SFT followed by RL improves the 4B manager",
|
||
"quote": "SFT followed by RL improves the 4B manager’s JJ score from 76.58% to 88.80%.\nUni-Mem also outperforms all off-the-shelf managers, including the strong",
|
||
"corpus_sha256": "b83a772fb6a1340b2440359bec2bf3752d583d86e213806c5dbf7a648d947bdd"
|
||
},
|
||
{
|
||
"id": "U9",
|
||
"paper": "2610.29671",
|
||
"theme": "T2",
|
||
"claim": "Uni-Mem outperforms all off-the-shelf managers including Qwen3.5-27B (87.05%).",
|
||
"value": "87.05%",
|
||
"note": "Scale is not the binding constraint; joint training is.",
|
||
"anchor": "Uni-Mem also outperforms all off-the-shelf managers",
|
||
"quote": "Uni-Mem also outperforms all off-the-shelf managers, including the strongest one, Qwen3.5-27B (87.05%), and the largest one, Qwen3.5-397B-A17B (85.86%).\nThese results demonstrate that a small manager trained jo",
|
||
"corpus_sha256": "b83a772fb6a1340b2440359bec2bf3752d583d86e213806c5dbf7a648d947bdd"
|
||
}
|
||
]
|
||
}
|