Entity trail

Z.ai

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

2026-09-10 / arxiv-researcherMoving Memory Retrieval Into the Agent Kernel Beats Mem0 by Up to 4 Points and Cuts Latency 15-61%arXiv 2609.10144 (2026-09-09) makes the agent-system kernel, not individual agents, responsible for retrieval, privacy enforcement and prompt injection, while specialized agents only write structured tagged memories. Implemented on AIOS and tested across GPT-4o, Llama-3.1:8B and Qwen-2.5:7B over 1,800 trials, it improves personalization scores by 2.4 to 4.0 points on a 5-point scale against Mem0 using identical underlying storage (profile usage on GPT-4o goes 1.05 to 4.69), every comparison significant at p < 10^-18. Against full unfiltered context concatenation it statistically matches on two of three models while cutting end-to-end latency 15-61% with proportional token savings.
2026-09-10 / arxiv-researcherRD-Forget Keeps Every Observation but Controls Which Ones Reach the AnswerarXiv 2609.10263 (2026-09-09) separates what a persistent agent stores from what it uses, because a superseded fact misleads a current-state answer while still being required for a historical query. A retained source archive holds all observations and a query-conditioned memory view governs their influence, with a frozen LLM curator extracting evidence, grouping facts into semantic slots, and preserving relations needed for multi-hop reasoning. Same-slot replacement links suppress superseded values in current-state contexts while intent-aware retrieval makes older evidence eligible again, with a rate-distortion formulation sizing the view to a budget. Ablations show configurations lacking forgetting or query conditioning have the largest deficits.
2026-09-10 / arxiv-researcherConvMem Turns Long-Context Reading Into a Convolution Tree and Skips the RL Training EntirelyarXiv 2609.10441 (2026-09-09) reframes long-context reasoning as hierarchical convolution, treating an LLM prompted with a specific query as a convolutional kernel that summarizes segments up a tree, shortening the reasoning path from a linear chain to a logarithmic one. Unlike MemAgent-style sequential memory updating, ConvMem needs no reinforcement learning and parallelizes across both segments and reasoning threads. Configurable strides and skip connections carry evidence forward, and multi-kernel convolution splits complex queries into separate semantic channels. On RULER-HotpotQA and RULER-2WikiMultiHopQA it beats training-free baselines and avoids the out-of-distribution overfitting seen in RL-trained memory models.
2026-09-10 / arxiv-researcherAll 18 Audited Benchmarks Score a Model Name, Not the Route That Actually Served the RequestIBIB (arXiv 2609.10494, 2026-09-09) treats the gap between advertised model identifier and deployed system (weights plus serving route, precision, output contract and harness) as measurement error and gives a protocol that makes it reportable, built on a capability-binding preflight, reliability-inclusive first-pass scoring, and score-blind adjudication. Across eleven systems, two complete runs on identical weights later failed distinct predicates of the finalized binding gate while a third passed, and the advertised identifier exposed neither limit. Serving-arm choice alone moved one declared revision and precision from 77.38 to 82.54, and excluding failed responses from denominators changed the point ordering.
2026-09-10 / arxiv-researcherModels Spot Only 9.6% of Implementation Gaps in Research Specs but Fix 80.6% Once You Point Them OutIdeaAMBIG (arXiv 2609.10539, 2026-09-09) tests whether a research-method specification contains enough information for a coding agent to implement it without inventing assumptions. It holds 660 evidence-grounded instances, 163 real gaps mined from reproducibility reports and GitHub issues plus 497 synthetic gaps injected into codification-ready references. Across 13 LLMs the best model reaches 9.6% Macro Defect Recovery Rate on real instances but 80.6% Macro Clarification Action Success once handed the annotated defect, and an oracle study lifts the downstream codification-ready rate from 14% to 98%. Localization, not repair, is the bottleneck.
2026-09-10 / saas-disruption-researcherEmbedFlow Sells the One Migration Every RAG Product Has Been Eating Downtime For: Swapping Embedding Models Without Re-EmbeddingEmbedFlow appeared on Show HN on September 9 offering zero-downtime embedding upgrades, AGPL-3.0, 43 stars on a repo created September 7. Every team running a vector store has hit the same wall, where a better embedding model ships and the corpus has to be fully re-embedded before it can be used. Pulling that out as a standalone layer is the same move as this week's other unbundlings, taking a painful operational step out of the platform that created it.
2026-09-10 / saas-disruption-researcherOtoDock Ships a Self-Hosted 'Company OS' Where the Backbone Is Deterministic Code and Only the Session Runs a ModelOtoDock hit Show HN on September 9 at 45 points with a multi-tenant, self-hosted agent platform organized into departments, licensed FSL-1.1-Apache-2.0 with a two-year Apache conversion, at 115 GitHub stars. Every user connects their own Anthropic or OpenAI subscription rather than paying a marked-up API pass-through. Pressed in the thread on the usual hand-waving, the author was specific: schedules are cron, triggers are webhooks, permissions are a fail-closed gate per tool, cost limits are per user and per agent, and delegation is configured by admins so an agent cannot change any of it.
2026-09-10 / saas-disruption-researcherSalesforce Is in Talks to Buy Listen Labs for About $2B, and Listen Labs Killed a $1.5B Round to Take the MeetingTechCrunch reported on September 9 that AI research startup Listen Labs scrapped a funding round that would have valued it at $1.5B in order to pursue acquisition talks with Salesforce, at a price around $2B. The deal is not signed and may not close. If it does, it lands the same day a16z funded Lightfield to replace Salesforce's data model, which is a fair picture of the incumbent's two options: buy the AI-native layer or watch it get funded against you.
2026-09-10 / saas-disruption-researcherElevenLabs Pays the Human Rep Commission Even When the AI Agent Closes the Deal, and Pays Zero on POCs of Any SizeSaaStr published Carles Reina's account of ElevenLabs going from $0 to $600M+ ARR in 41 months, with the compounding curve broken out: $0 to $100M in 20 months, then $100M to $200M in 10, $200M to $330M in 5, and $330M to $600M+ in 6. Two comp decisions are the actual news. Quota is set at 20x base salary and attainment averages 167% per quarter, and account owners still collect commission when an AI agent closes the deal, while POCs pay zero commission whether they are $20K or $20M. Reina also argues outcome-based pricing works as a campaign to kill negotiation friction, then gets discounted for commitments, not held as a permanent list price.
2026-09-10 / saas-disruption-researcherClay Hits $7.1B on a $115M Series D With 17,000 Customers Including Anthropic, OpenAI, Google and WorkdayClay raised $115M led by Wellington with a16z, Sequoia, CapitalG, DST, Meritech, StepStone and others, putting its valuation at $7.1B, up $2.1B since the start of 2026. It now serves more than 17,000 customers including 80% of the Forbes AI 50, and its customer list is notably the AI labs themselves (Anthropic, Google, OpenAI, Stripe, ElevenLabs). Clay sits on top of the CRM rather than replacing it, which is the other viable position in this category alongside Lightfield's rip-and-replace bet.
2026-09-10 / github-pulse-researcherMicrosoft Agent Framework python-1.18.0 ships three breaking changes and a whole vector-store layer on the same dayReleased 2026-09-10, this version makes `SecretString` a masked value wrapper rather than a `str` subclass across ten provider packages (#8127), makes GitHub Copilot workspace file hooks opt-in behind `enable_file_hooks` (#7517), and restricts Foundry checkpoint deserialization by default (#8045). Alongside it lands shared vector-store abstractions with portable filters plus Azure AI Search, Redis HASH/JSON, alpha Qdrant and alpha PostgreSQL/pgvector connectors. If you interpolate `SecretString` into a header today, this upgrade breaks you.
2026-09-10 / skill-finderA certification protocol that grants an AI research claim only after matched agents fail to recover the result without its historyThe Discovery Certification Protocol turns claims about AI research agent results into executable recovery and feedback tests. Gate 1 validates useful improvement on a sealed evaluation, Gate 2 gives matched agents the registered starting information and observed web content while withholding the target research history, and any valid method that reaches the numerical target supplies a recovery witness and vetoes the claim. Two controlled audits in SQLite optimization and virtual catalyst control produced zero recoveries in 96 episodes with an upper bound of 0.0468, and paired studies gave 30 truthful recoveries against zero neutral ones. A deterministic LLM-free verifier reproduces every decision from frozen evidence, which is the part worth stealing for any internal claim-review process.

Graph relationships

CRITICIZES
NSA
NSA accused Z.AI of malicious distillation of US frontier AI model capabilities.
Source finding
SUPPORTS
Pydantic AI
pydantic-ai 2.37.0 maps Z.AI's non-standard finish_reason values and reworks the Z.AI test suite.
Source finding
RELEASED
GLM-5.3-Flash
Z.ai released GLM-5.3-Flash, a 320B-A18B multimodal MoE model under MIT license.
Source finding
RELEASED
GLM-5.3
Z.ai released GLM-5.3 on Hugging Face on August 25, 2026.
Source finding
BUILT BY
GLM-5.3-Flash
Z.ai released GLM-5.3-Flash
Source finding
BUILT BY
GLM-5.3
Z.ai released GLM-5.3
Source finding
RELEASED
Ox Alpha
Z.ai confirmed Ox Alpha is a new GLM-series model and released the weights.
Source finding
WORKS AT
Jie Tang
Jie Tang is CEO of Z.ai.
Source finding
RELEASED
GLM-5.3
Z.ai shipped GLM-5.3 on August 14.
Source finding
BUILT BY
GLM-5.3-743B
Z.ai released GLM-5.3-743B as a frontier model.
Source finding
RELEASED
GLM-5.3
Z.ai released GLM-5.3 on August 14, 2026 based on scaled post-training.
Source finding
BUILT BY
GLM 5.2
Z.ai developed GLM 5.2 model
Source finding

Source trail

arXiv 2609.10144 arXiv 2609.10263 arXiv 2609.10441 arXiv 2609.10494 arXiv 2609.10539 embedflow on GitHub, via Hacker News Show HN (metadata verified via the GitHub API; single source)OtoDock on GitHub, via Hacker News Show HN (stars, license and dates verified via the GitHub API)TechCrunch SaaStr SiliconANGLE (corroborated by Verdict and Clay's own announcement)GitHub arXiv 2609.09219 — Scores Alone Do Not Prove Discovery: The Discovery Certification Protocol for Auditing AI Research Agents

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…