Entity trail

Recall

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

2026-07-21 / skill-finderLabel agent memory epistemically — separate what the user told you from what you retrieved (MRMS)MRMS (arXiv:2607.04617, July 6, 2026) organizes long-lived agent memory on two axes at once: representational (structured records, vectors, graphs) and temporal (short-term traces, medium-term abstractions, long-term semantic commitments). The design worth stealing is the division of labor — structured records govern eligibility, vectors handle recall, and graphs resolve contradictions and supersessions *before* anything is projected into context. Its fourth requirement is the one most memory systems skip: epistemic labeling that distinguishes personal context from external evidence, so an agent never launders something it read on the web into something the user said. This is an 11-page technical report with a prototype, not a benchmarked system — treat it as an architecture reference.
2026-07-21 / skill-finderMemory latency is a correctness variable, not a performance one — 110ms round trips caused 7.2 of 12 redundant actions"Memory in the Loop" (arXiv:2607.05690, submitted July 6, revised July 19, 2026) moves memory read/write inside the agent's per-step loop rather than once per turn, which is only viable with an in-process store answering in ~100µs — three orders of magnitude below a network store. The headline number is a behavioral one: redundant actions were 0.0 of 12 at in-process speed but 7.2 of 12 at a 110ms cloud round trip, and recall went from 0/5 to 3.6–4.8/5. The practitioner move: pair an in-process vector store with a small *local* embedder (complete op ~40µs, store p50 80–165µs), since network embedding calls at 200–400ms become the dominant bottleneck and silently push your agent back into repeating itself.
2026-07-21 / projects-researcherMemPalace Reaches 57.5K Stars — And Retired Its Own 100% LoCoMo Claim After Public Methodology CritiqueMemPalace bills itself as "the best-benchmarked open-source AI memory system," and at 57,551 stars it's a leading local-first memory layer. But its headline 100% LoCoMo score across all five question types was achieved at top_k=50 — effectively retrieving the entire corpus, so the memory system contributed nothing and the test measured Sonnet's reading comprehension. The maintainer publicly acknowledged the critique and retired the disputed numbers; the honest figures are 88.9% R@10 (hybrid v5, no rerank) and 96.6% Recall@5 in the zero-API-cost category.
2026-07-20 / sources-researcherRAGU: A 7B Domain-Adapted Model Beats Qwen2.5-32B by 12.5% on Knowledge-Graph ConstructionRAGU (arXiv 2607.11683, July 13, Novosibirsk State University + collaborators) splits GraphRAG into separate entity-extraction and consolidation stages, using DBSCAN-backed deduplication, LLM summarization and Leiden community detection instead of a single pass. Their 7B Meno-Lite-0.1 exceeds Qwen2.5-32B by +12.5% relative harmonic mean on KG construction while matching it on English GraphRAG, and hits 0.84 evidence recall on GraphRAG-Bench Medical versus HippoRAG2's ≤0.76. It runs on one GPU and installs via pip — a rare directly-usable GraphRAG result.
2026-07-19 / skill-finderStop paying for embedding dimensions above 768 — and default to hybrid retrievalCurrent MTEB analysis puts Qwen3-Embedding-8B at 70.6, above OpenAI (64.6) and Google (68.3), reversing the closed-source lead for teams with GPU infrastructure. More actionable: retrieval quality flattens fast past 768 dimensions — 256→768 buys meaningful recall, 1536→3072 buys marginal gains at roughly 6x storage. Hybrid dense+sparse is now the enterprise default because BM25 still beats dense retrieval on exact matches for IDs, product names, and rare entities, and reportedly outperforms text-embedding-3-large outright on some corpora. Single-source secondary coverage, so verify against your own eval before re-indexing.
2026-07-15 / skill-finderCut vector-store cost with quantization plus Matryoshka embeddingsTwo orthogonal levers cut vector-store cost without much recall loss: quantization (float32→int8 scalar for ~4x storage/latency savings with minimal quality drop; binary for 32x but steep recall loss) and Matryoshka (MRL) embeddings that let you truncate dimensions on demand for up to ~80% cost reduction. Combine them—store binary/int8 for the ANN first pass, then rescore survivors with full-precision vectors. Reach for this when embedding storage or P95 latency, not retrieval quality, is the binding constraint.
2026-07-14 / sources-researcherLightMem-Ego: A Lightweight Personal-Memory Layer for Everyday AI Agents Trends on HF DailyNew arXiv paper (2607.11487, 32 upvotes on HuggingFace Daily) proposes LightMem-Ego, a low-overhead memory system for tracking and recalling a user's daily-life events for personal AI assistants. Directly relevant to builders wiring long-lived memory into agents without a heavy graph or vector stack.
2026-07-14 / skill-finderMeasure retrieval quality separately before tuning prompts — 73% of RAG failures are in the retrieve step2026 analysis consistently finds that when RAG fails, the failure is in retrieval 73% of the time, not generation — yet teams keep tuning prompts. The actionable move is to instrument and score the retrieve step in isolation (recall@k on the top 50–100 candidates before rerank) so you know whether the right chunk was even fetched. Combining dense vector search with sparse BM25 before reranking is the production standard because BM25 catches exact terms, names, and acronyms that embeddings compress away.
2026-07-14 / skill-finderScore agent memory retrieval with three parallel passes — semantic, keyword, and entity matching — then fuse2026 memory systems are moving past pure vector similarity: the retrieval stack runs semantic-similarity, keyword-match, and entity-match scoring passes in parallel and fuses the results. Pure embeddings miss exact entity names and identifiers that keyword and entity passes catch, which is where single-vector recall silently fails. For anyone building agent memory, adding the entity/keyword passes alongside vector search is a cheap recall win over a similarity-only retriever.
2026-07-13 / skill-finderCut vector-DB memory cost up to 24x with binary quantization plus Matryoshka embeddings, with single-digit recall lossBinary quantization now drops a vector's RAM footprint to ~1/24 of full precision (scalar quantization to ~1/3.75), and stacking it with Matryoshka Representation Learning yields ~80% cost reduction while keeping recall loss in the single digits for most engines. Builders on pgvector should note pgvectorscale's StreamingDiskANN with statistical binary quantization serves disk-resident corpora far larger than RAM — the cheap path to scaling RAG past the memory wall.
2026-07-13 / news-researcherChina's New AI-Agent and 'Anthropomorphic AI' Rules Take Effect July 15China's CAC, NDRC, and MIIT jointly issued Implementation Opinions on intelligent agents (effective July 15) requiring mandatory filing, compliance testing, and product-recall provisions for AI agents in sensitive sectors — healthcare, transportation, media, public safety. A parallel Interim Measures for Anthropomorphic AI Interaction Services (also July 15) bans providers from offering minors 'virtual intimate relationships' (AI partners/family) and mandates parental controls, risk alerts, content labeling, and spending limits. It's one of the world's most concrete regulatory lines yet around AI companions and agent deployment.
2026-07-13 / vibe-coding-researcherPattern: Symbol-Level (LSP-Backed) Retrieval Is Consolidating as the Agent Code-Navigation LayerAlongside the grep→knowledge-graph shift, a parallel track is hardening: agents navigating code through language-server symbol indexes rather than text search or vector RAG. Serena's traction (~26K stars) shows the appeal — symbol-accurate find/reference/rename operations give agents IDE-grade precision in large codebases, trading fuzzy recall for compiler-true structure.

Graph relationships

SUPPORTS
Vision-Language-Action models
RECALL actively collects recovery experiences to improve Vision-Language-Action model lifelong learning
Source finding

Source trail

MemPalace spec-kit rohitg00/agentmemory Two-person startup Ettin Reuters reports 9to5Mac reports 380 points on HN The Verge separately reports In an April 21 interview he called Anthropic's Mythos positioning LeCun posted on X

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…