AuthorsNeeraj Yadav
TL;DR
MemStrata uses deterministic supersession over a bi temporal ledger to cut RAG’s stale fact errors from 36–38% to ≈0 on real GitHub fixes.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
RAG retrieves both old and new code facts with near identical similarity, so it cannot tell which value is current and often serves stale answers.
On SWE bench GitHub buggy to fixed histories, naive_rag and advanced_rag commit the superseded value in 36.1–37.7 percent of forced answers, breaking coding assistants’ reliability.
HOW IT WORKS
MemStrata routes each value bearing turn through a deterministic assertion path backed by a production triple extractor and a bi temporal ledger of assertions.
Like a card catalog that always files the latest card on top, MemStrata closes old validity intervals and opens new ones whenever a clean subject relation object triple changes.
This strict_object_supersede rule lets MemStrata retire stale values and surface only currently valid rows, something a plain similarity based context window cannot guarantee.
DIAGRAM
This diagram shows how MemStrata constructs marker free longitudinal scenarios from SWE bench buggy to fixed GitHub issues.
DIAGRAM
This diagram shows how MemStrata and RAG are evaluated in paired allowed and forced regimes on the same 130 cached scenarios.
PROCESS
Constructing real longitudinal scenarios
MemStrata uses a local LLM patch reader and the production triple extractor to emit a single atomic subject state_a state_b question tuple from a SWE bench issue.
Deterministic guard
MemStrata applies a deterministic guard that rejects scenarios unless the two values differ, are atomic, and carry no recency tells, ensuring clean supersession engagement.
Self validation selection criterion
MemStrata keeps a scenario only if the production triple extractor keys the state A and state B sentences identically with objects equal to the two values.
Deterministic supersession and retrieval
MemStrata ingests state A then state B into the bi temporal ledger via the deterministic assertion path and strict_object_supersede, then answers the question by retrieving only currently valid rows.
KEY CONTRIBUTIONS
A real data longitudinal benchmark
MemStrata mines 130 clean atomic state transitions from 707 SWE bench Lite plus Verified GitHub issues, with a marker free invariant and explicit selection via the production triple extractor.
An end to end win on real history
MemStrata achieves 0.908 allowed and 0.985 forced accuracy, while naive_rag and advanced_rag reach only 0.569–0.615, and MemStrata drives stale fact error from 0.361–0.377 down to ≈0.
Moat preserving bug fix in strict_object_supersede
MemStrata identifies a case and punctuation insensitive comparison bug in the assertion path, fixes it with strict_object_supersede, and verifies synthetic accuracy at 1.000 and stale fact error at 0.000.
RESULTS
accuracy — allowed
0.908
+0.339 over naive_rag
accuracy — forced
0.985
+0.370 over naive_rag
stale-fact-error — forced
≈0.00
36.1–37.7 points lower than naive_rag and advanced_rag
mean retrieval latency
2.13 s
similar to naive_rag’s 2.16 s and 8.5 times faster than advanced_rag’s 18.1 s
On 130 SWE bench Lite plus Verified atomic transitions, MemStrata is evaluated against naive_rag and advanced_rag in paired allowed and forced regimes. The MAIN_RESULT shows MemStrata maintains current code facts with near zero stale commitments while matching RAG latency and compressing active memory by 48 percent.
BENCHMARK
On 130 SWE bench Lite plus Verified atomic transitions, MemStrata is evaluated against naive_rag and advanced_rag in paired allowed and forced regimes. The MAIN_RESULT shows MemStrata maintains current code facts with near zero stale commitments while matching RAG latency and compressing active memory by 48 percent.
BENCHMARK
Accuracy — allowed on SWE bench Lite plus Verified atomic transitions.
KEY INSIGHT
advanced_rag, which adds an LLM reranker, has stale fact error 0.377 when forced, slightly worse than naive_rag’s 0.361 on the same scenarios.
Designers often expect smarter reranking to help, but without a temporal signal advanced_rag simply reshuffles chunks and cannot distinguish stale from current values, sometimes amplifying stale commitments.
WHY IT MATTERS
MemStrata shows that deterministic supersession over a bi temporal ledger can keep current code facts across evolving GitHub histories at RAG latency.
Builders can now ship coding assistants whose session memory has a moat like guarantee against stale fact errors on clean code mutations, rather than hoping embedding similarity and reranking will suffice.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Okan Bursa
· 2026
Adaptive RAG Memory (ARM) augments a standard retriever–generator stack with a Dynamic Embedding Layer and Remembrance Engine that track usage statistics and apply selective remembrance and decay to embeddings. On a lightweight retrieval benchmark, ARM achieves NDCG@5 ≈ 0.9401 and Recall@5 = 1.000 with 22M parameters, matching larger baselines like gte-small while providing the best efficiency among ultra-efficient models.
Zhanghao Hu, Qinglin Zhu et al.
· 2026
xMemory organises agent memory via segments, memory components, groups, and dynamic maintenance with a sparsity plus semantic faithfulness objective. On LoCoMo with GPT-5 nano, xMemory reaches 38.71 BLEU and 50.00 F1 versus Nemori’s 36.65 BLEU and 48.17 F1, while cutting tokens per query from 9154.76 to 6581.20.
Answers use this explainer on Memory Papers.
Checking…