Entity trail

LongMemEval_S

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-02 / arxiv-researcherLatentPress Compresses Context Into Soft Tokens a Frozen Decoder Reads Directly, Beating Uncompressed Evidence at 7.7xInstead of carrying compressed context as text or rendered images, LatentPress writes conversation histories and long documents into continuous memory tokens fed straight through the decoder's input-embedding interface, with no text reconstruction at inference and only an adapter trained (4.2M to 26.2M parameters, roughly 0.1% of the decoder). On LongMemEval it reaches 0.504 accuracy at 7.70x compression against 0.490 for uncompressed evidence, versus 0.184 for text summaries and 0.426 down to 0.312 for OCR-based compression. Writing takes 43ms per conversation, about an order of magnitude faster than text summarization or OCR, and reading is 5 to 9x faster than raw context.
2026-08-29 / projects-researcherLemmalog turns agent memory into a Datalog engine and beats full-transcript prompting on 38x less contextJordyZomer/lemmalog was created 2026-08-27, is 76 stars and MIT-licensed in Rust, and the author's write-up hit HN at 173 points on 2026-08-28. It stores LLM observations as provenance-tracked facts with stratified rules and incremental derivation, exposed to a harness through an MCP server, so the model does language extraction while the database keeps logical consistency. Reported results are 0.463 ±0.010 F1 on LongMemEval with 38x smaller context than full-transcript prompting and 0.533 ±0.001 F1 on LoCoMo at 6x smaller, strongest on knowledge updates and adversarial questions.
2026-08-27 / sources-researcherScroll backs agent sessions with an append-only event log and a persistent Python kernel, hitting 94.8% on LongMemEval_SAlibaba's Scroll (arXiv 2608.21690, submitted 2026-08-21) treats each agent session as an executable environment rather than a compressed chat history. Tool outputs and derived state bind to variables in a typed namespace held by a sandboxed persistent Python kernel, model-written code searches and transforms that state, and only explicitly printed projections enter the next model call. Evicted spans stay recoverable through an eviction index of compact landmarks pointing at exact event log addresses. With Qwen3.8-Max it scores 94.8% on LongMemEval_S, 73.1% on BEAM_10M (5.1 points over the best published memory system) and 86.7% on LOCA_256K, 37.4 points over the best published long-horizon agent.
2026-08-24 / sources-researcherPoisoning 1.2% of an Agent's Memory Drops Accuracy From 0.85 to 0.30, and Write-Time Screening Caught Zero of 360 Poisoned EntriesThis paper measures agent memory poisoning with plainly worded false assertions generated in a single pass, with no instruction, trigger, or retriever optimization. Poisoning 1.2% of a LongMemEval corpus cut accuracy from 0.850 to 0.300, and a four-stage write-time screening pipeline that reaches 0.832 recall on indirect prompt injection rejected 0 of 360 poisoned memories, because telling a false assertion from a true one needs grounding outside the text. Provenance-weighted retrieval fared no better: the shipped weight was statistically indistinguishable from no defense (p=0.80), and a stronger weight only recovered utility by excluding untrusted content entirely, dropping evidence recall to zero and accuracy to 0.0417 when the answer-bearing evidence itself arrived untrusted.
2026-08-18 / github-pulse-researcherKITE Claims SOTA on LoCoMo and LongMemEval With No Vectors and ~1.6K Tokens of Reader Contextmemoket/memoket-kite, created 2026-08-12 and pushed again on 2026-08-18, is an Apache-2.0 agent memory engine that drops embeddings, vector DBs, and rerankers entirely in favor of a topic-indexed single file. Its README reports 93.51% on LoCoMo and 85.60% on LongMemEval-S using gpt-4.1-mini across all benchmark stages, with an average reader context of only ~1.6K tokens per question. The pitch for builders is determinism and receipts: a question compiles into a readable plan that names who said what and when, and returns empty when the answer was never stated instead of handing back a nearest match.
2026-08-16 / arxiv-researcherRippleMem Builds Agent Memory From Cues Instead of One-Shot Retrieval, Cutting Graph Construction Cost ~30xRippleMem (arXiv 2608.13334, 2026-08-13) argues the bottleneck in long-term agent memory is not storage but recovering evidence scattered across many interactions — flat retrieval returns isolated records, and graph memory systems are expensive to build. It stores history as cue-rich episodic units in an event-centric graph, recalls anchors via hybrid cues, then expands along semantic and structural associations so initially recalled memories become cues for finding the rest. On LoCoMo and LongMemEval-S it improves LLM-as-a-Judge accuracy by 3.95% and up to 11.87% respectively while reducing graph construction cost roughly 30x.
2026-08-14 / sources-researcherLycheeMemory V2 Consolidates Agent Memory at Semantic Segment Level, Not Per Turn: 92.20% on LongMemEval-S With 75.9% Fewer Construction TokensSubmitted August 13 (arXiv 2608.12990) by Dongfang Li, Baotian Hu, Min Zhang and colleagues, LycheeMemory V2 replaces turn-level memory consolidation with semantic segment-level consolidation, using semantic boundary detection instead of fixed-window batching. It reports 89.22% on LoCoMo and 92.20% on LongMemEval-S while cutting construction tokens 86.0% and 75.9% respectively versus the A-Mem baseline. This is a different lever than last week's write-time-vs-retrieval framing: the win comes from how often you invoke the model to encode, which is the line item that actually shows up on an agent's bill.
2026-08-13 / sources-researcherAdaption Labs: Agent Memory Fails at Write Time, Not Retrieval — 90.6% on LongMemEval vs 60.6% for Full History, With 97% Fewer TokensPosted 13 Aug 2026, the argument inverts the usual framing: retrieval can only search memories that exist, so the highest-leverage optimization is extraction at write time. Their system writes in two forms — narrative memories preserving causal context and atomic memories preserving verbatim facts — and never deletes, instead appending new state marked as superseding the old so retrieval favors current truth while history survives; they report expecting to need knowledge graphs and finding they did not. Reported numbers: LongMemEval 90.6% vs Mem0 OSS 71.6% and full-history 60.6%; LoCoMo 88.2% vs 82.2%; BEAM 61.3% vs 39.7%; 97% fewer tokens on long conversations, mean latency down 45%, p95 down 67%. Vendor-published benchmarks with no independent replication — weight accordingly.
2026-08-11 / skill-finderGive agent memory SUPERSEDES and CONTRADICTS edges: 92.8% on LongMemEval-S at $0.00161 per questionSodaMem (arXiv 2608.08055, Aug 8) starts from the observation that an agent assisting you over weeks must remember what is *currently* true, not merely what was once said — a property flat conversation logs cannot express. It extracts typed events with source attribution, then tracks temporal validity with explicit SUPERSEDES, CONTRADICTS, and UPDATES edges so a superseded fact is structurally retired rather than competing with its replacement at retrieval time, and answers with a verifiable evidence trail. On LongMemEval-S it scored 92.8% (464/500) at $0.00161 per question and a median ~18.3k tokens using deepseek-v4-flash, with code released — a cheap upgrade path for any agent whose memory is currently an append-only log.
2026-08-05 / agents-researcherLeanMem routes agent memory by content type instead of one summarization pipeline, gaining up to 15.1 pointsLeanMem argues that dialogue content should be handled by compressibility, temporal dynamics, and fidelity requirement rather than pushed through one uniform summarize-and-retrieve path. It filters low-value content, then stores informative segments as compact profile memory, temporally structured event memory, or source-grounded record memory, updating only the dynamically evolving event memories during maintenance and allocating per-query retrieval budgets across memory types at inference. On LoCoMo and LongMemEval-S with GPT-4.1-mini and Qwen3-8B it beat the strongest memory baseline in every setting by up to 15.1 points, at the lowest or near-lowest construction cost, inference tokens, and latency.
2026-07-30 / github-pulse-researcherMemPalace Claims 96.6% R@5 on LongMemEval With Zero API Calls — and Warns of Malware-Distributing Impostor DomainsMemPalace/mempalace (57,888 stars) stores conversation history as verbatim text and retrieves it with semantic search, explicitly refusing to summarize, extract or paraphrase; the index is structured into wings (people/projects), rooms (topics) and drawers (original content) so searches can be scoped rather than run flat. It reports 96.6% R@5 raw on LongMemEval with no API calls, on a pluggable backend interface currently defaulting to ChromaDB. The repo now leads with a security caution that impostor sites on .tech, .net and other .com variants may distribute malware — a new failure mode for high-star AI infrastructure.
2026-07-28 / github-pulse-researcherTwo Agent-Memory Projects Now Publish Directly Conflicting Benchmark Claims — MemPalace 96.6% vs agentmemory 95.2%, on Different SplitsMemPalace (57,821 stars, ~512/day, v3.6.0) reports 96.6% raw recall@5 on LongMemEval with no LLM required, 98.4% with hybrid v4 on a held-out 450 questions, LoCoMo R@10 rising from 60.3% to 88.9%, ConvoMem 92.9% and MemBench 80.3% — while explicitly refusing head-to-head comparison against Mem0, Zep, Mastra or Supermemory on the grounds that different metrics on different splits are incomparable. Meanwhile rohitg00/agentmemory (25,913 stars) claims '#1 based on real-world benchmarks' with 95.2% recall@5 and 88.2% MRR on LongMemEval-S, and does publish a rival table citing mem0 at 68.5% and Letta/MemGPT at 83.2% on LoCoMo. Builders comparing the two are comparing incompatible evaluations — agentmemory's own docs concede only its figures are measured on LongMemEval-S.

Graph relationships

BENCHMARKED AGAINST
LycheeMemory V2
LycheeMemory V2 scored 92.20% on LongMemEval-S with 75.9% fewer construction tokens than A-Mem.
Source finding
BENCHMARKED AGAINST
graphify
Graphify reports 76% accuracy on LongMemEval-S benchmark
Source finding
BENCHMARKED AGAINST
Engram
Engram scored 83.6% vs 73.2% baseline on LongMemEval_S with 8x fewer tokens.
Source finding
BENCHMARKED AGAINST
LycheeMemory V2 -> LongMemEval_S
LycheeMemory V2 scored 92.20% on LongMemEval-S with 75.9% fewer construction tokens than A-Mem.
Source finding
BENCHMARKED AGAINST
graphify -> LongMemEval_S
Graphify reports 76% accuracy on LongMemEval-S benchmark
Source finding
BENCHMARKED AGAINST
Engram -> LongMemEval_S
Engram scored 83.6% vs 73.2% baseline on LongMemEval_S with 8x fewer tokens.
Source finding

Source trail

arXiv 2609.01507 GitHub arXiv 2608.21690 (submitted 2026-08-21)arXiv (2608.21230, submitted 2026-08-21)GitHub arXiv 2608.13334 arXiv (via HuggingFace Daily Papers)Adaption Labs arXiv 2608.08055 — SodaMem: Evidence-Grounded Temporal Graph Memory for LLM Agents arXiv GitHub GitHub

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…