An episode-based AI memory system, TSIM, recorded the highest accuracy in all three reported 128k answer-backend blocks and scored higher than native Full Context in separate diagnostics. The results came from SCALE-QA, a synthetic, counterfactually constructed benchmark of flat, unsegmented, mixed-topic conversation threads in which a later task question depends on evidence from earlier in the exchange.
TSIM uses semantic-shift episode segmentation, multi-view episode indexing and evidence-first episode ranking before assembling compact evidence for the answer model. In practice, it identifies the relevant episode first and gives the answer model a compact set of evidence.
TSIM led across three answer models
The main comparison used three answer backends at 128k tokens. It compared TSIM with Standard RAG, Hybrid-RRF Chunk RAG, RAPTOR, MemGPT and HippoRAG. Full Context was included as a native-context diagnostic.
At that setting, TSIM reached 69.6% accuracy with Gemma2:9b, 80.2% with Gemini 2.5 Flash and 73.8% with GPT-4o-mini, the highest result in each backend block.
A tuned non-episodic control reached 56.2% accuracy and 63.5% on CL Hit, the study's closed-loop expected-evidence measure, 17.6 percentage points below TSIM. With the Gemini backend, the reported TSIM-MemGPT gap was 5.58 points, with a paired-bootstrap 95% confidence interval of 4.15 to 6.97 points.
The comparison used a shared runtime protocol. Systems received identical packages, seeds, noise and batch mappings, and evaluation reset between packages. Under that protocol, the intended comparison was the memory or retrieval strategy.
Experiments used all 3,000 questions through 128k, while the 1M analysis used a stratified 400-question subset.
More context did not guarantee better scores
A reported ablation under GPT-4o-mini showed rising scores across the tested variants: 26.2% for Standard RAG, 43.4% for fixed-token direct retrieval, 55.5% for semantic-drift episodes and 74.2% for the full multi-view episode memory stack.
The same distinction appeared as the context window grew. Under GPT-4o-mini, Full Context fell from 62.5% at 16k to 29.8% at 128k, whereas TSIM remained at 73.8% using about 1k retrieved tokens.
On a separate stratified 400-question diagnostic, DeepSeek R1 with Full Context reached 81.2% at 128k, versus 93.8% for TSIM. In a 1M-token Gemini 2.5 Flash diagnostic, Full Context reached 87.2%, using 1.05M prompt tokens and 23.87 seconds, while TSIM reached 96.5% with about 1.3k retrieved tokens and 2.16 seconds.
Those token and timing figures describe the reported implementation and hardware, rather than a hardware-normalized or universal cost comparison.
The retrieval result
The retrieval diagnostics measured the same distinction. Across the full 3,000-question run, TSIM's episode retrieval recovered all required evidence among its top five results in 81.0% of cases and kept that evidence together in 89.0%. Standard RAG chunks did so in 45.6% and 0.4% of cases, respectively.
An additional transfer check used 500 cleaned V1 questions from LongMemEval-S. TSIM reached 71.0% judged accuracy, compared with 61.2% for BM25 fixed-chunk retrieval and 56.6% for BGE turn retrieval. The paper describes this as a transductive diagnostic, not a held-out estimate of generalization.
One comparison was less decisive. Against a context-matched fixed-chunk control, TSIM was 9.8 points higher in accuracy and 18.30 points higher in all-evidence recall. Against BGE turn retrieval, its accuracy edge was 14.4 points. A boundary-assisted session diagnostic had a 3.4-point advantage over TSIM, but it used 2.77 times more answer context and the difference was not significant, with a 95% confidence interval for TSIM minus session of -7.2 to +0.6 points and McNemar p = 0.1109.
What the tests do and do not establish
SCALE-QA contains 3,000 QA records across 10 topics, with 300 questions per topic. Its validation report listed 4,346 audited evidence snippets, exact full-turn records for all 3,000 records and zero critical validation issues.
The benchmark's design shapes the meaning of the result. It tests whether a memory system can recover earlier evidence for a later task decision in flat, unsegmented, mixed-topic threads. The LongMemEval-S result is a transductive diagnostic rather than a held-out generalization estimate, so these findings concern benchmark accuracy, evidence recovery and the reported implementation, not a universal ranking of conversational memory systems.
The benchmark is synthetic and counterfactually constructed rather than sampled from naturally occurring assistant logs. The evaluation therefore does not establish how the system would perform on natural conversations.
The work is an arXiv preprint, version 1, dated 26 Aug. 2026. Formal study funding is not reported.
Paper data and sources
Original title: Reconstructing the Right Episode: Evaluating Interleaved Conversational Memory Beyond Long Context
Authors: Zhexi Feng, Ruiyi Zhang, Yongbo Yang, Pengtao Xie
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text