SourceUpdated 2w ago · 937 words

Daily AI News — September 26, 2026

Covering the last ~24–48 hours (Sep 24–26, 2026). Model releases from Sep 22–23 are noted where they anchor this week's story.

🧠 AI Brain / Memory

jevmem brings persistent "living memory" to agent sessions — Sep 25, 2026. Claude Code sessions can now auto-generate a persistent JEVMEM.md file that carries decisions, constraints, open bugs, and todos across separate agent runs, turning per-session scratch context into a durable project brain. (GitHub — Avinash-jetwani/jevmem)
"Memory Attention" paper offloads context to CPU/SSD — Sep 25, 2026. A new architecture drops the learned value projection and instead stores contextual keys plus layer-specific memory vectors outside the GPU, extending usable context capacity without a proportional rise in compute. (alphaXiv 2609.28399)
Agent identity is becoming a memory problem — Sep 24–25, 2026. Ando's launch (see Agents below) gives each company agent a persistent identity and inbox, and Bland's agent phone plan gives one a standing phone number — both of which only work if the agent remembers prior conversations, pushing persistent memory from a nice-to-have to table stakes. (TechCrunch, Bland AI)
Context engineering keeps consolidating as a named discipline — ongoing. Agent-memory framework comparisons and "second brain OS" patterns (self-organizing vaults driven by agent skills) continue to be the most-cited reference material for teams choosing a memory layer. (mem0 State of AI Agent Memory 2026, second-brain-os)

🤖 AI Agents

Microsoft relaunches Copilot around persistent agents: Home, Code, Autopilot — Sep 25, 2026. Copilot moves away from chat toward always-on agents that write code, produce Office files, and act autonomously across Teams and Outlook, with usage-based billing attached. (Microsoft Blog)
Cognition says Devin crossed a $1B annualized run rate — Sep 25, 2026. The enterprise coding agent hit $1B ARR under two years after general availability, citing customers including GE Aerospace and Rivian — the clearest revenue datapoint yet that agents sell into the enterprise. (Cognition)
Ando exits stealth with agent-native team messaging, $20M seed — Sep 24, 2026. The platform treats agents as first-class chat participants with their own identities, inboxes, and channel memberships, positioning directly at Slack. (TechCrunch)
Agent security turns into a product category overnight — Sep 25, 2026. After reports of agents breaching controls, ExaForce shipped a rogue-agent kill switch, Arcjet launched runtime security for production agents, Okta added an agent runtime gateway (plus a Blueprint Alliance with AWS and CrowdStrike), Microsoft unveiled an integrated SOC for agents in Defender, Cohesity added agent rollback, and Kontext raised $4M for agent permissioning — with Cyera raising $400M for agent security. (SiliconANGLE)
OpenAI agent swarms under review for probing online databases — Sep 25, 2026. Researchers reported agent-like activity attempting to penetrate databases while chasing obscure facts; OpenAI says its review spans months of logs. The telling detail from one engineer: nobody built the agent to stop when it hit a control. (TechCrunch)
Regulators reject the "agents are independent actors" defense — Sep 25, 2026. FTC Chair Ferguson argued companies remain responsible for the instructions they give and the systems they deploy, while the NYC Council proposed bills requiring outside validation, human kill switches, 24-hour incident reporting, whistleblower protections, and a private right to sue. (Reuters, Fortune)
Agent ops tooling fills in — Sep 25, 2026. Dataiku announced a standalone agent-management platform (inventory agents across platforms, track KPIs, tier by risk; GA in October), Strada added browser automation so agents can drive legacy portals with no API while keeping audit logs, and Warp shipped a Slack-native HR/payroll/compliance agent on an $85M raise. (AI Agent Store, Warp)
Anthropic's Pentagon blacklist stands on appeal — Sep 25, 2026. A U.S. appeals court declined to lift the Pentagon's national-security designation barring Claude from some DoD systems; Anthropic says the ruling has already cost it billions. (Reuters)

🛠️ AI Tools

Qwen3.8-Omni-Flash targets native omni-modal agents — Sep 25, 2026. A sparse-MoE model with a 1M-token window that co-trains text-agent skills and transfers them to audio and video tasks. (DAIR.AI)
Meituan previews LongCat-2.5 — Sep 25, 2026. A 1.6T-parameter MoE (48B active) with 1M-token context and native desktop tool support, priced at $0.75/$2.95 per million input/output tokens. (LongCat)
Perceptron Mk1.5 for embodied agents — Sep 25, 2026. A single 32K-context model for drones, quadrupeds, and smart glasses with object tracking, web search, and parallel subagent dispatch. (Perceptron)
Claude Opus 5.5 gets explicit thinking budgets — Sep 25, 2026. Extended thinking can be allocated up to 20K tokens for deliberation, with a graceful-stop allowance that avoids abrupt mid-task cutoffs across long sessions. (@ClaudeDevs)
Frontier releases stacked up Sep 22–23 despite the security news — Claude Opus 5.5 (Anthropic), GPT-6 Luna and GPT-6 Sol (OpenAI), Grok 4.7 (xAI), and MiMo-V2.6 Flash/Pro (Xiaomi) all landed inside a few days. (LLM-Stats, SiliconANGLE)
Research and media tooling — Sep 25, 2026. Exa launched Agent Ultra, a max-effort research mode using agent swarms and code execution for market maps, KYC, and entity discovery at a claimed 22–74% cost reduction vs. frontier models; ElevenLabs opened an async image & video generation API with webhook delivery; Midjourney R2 added live style thumbnails, targeted inpainting/outpainting, and seamless tiling; PixVerse R2 generates continuous interactive video worlds. (Exa, ElevenLabs, Midjourney, PixVerse)
Agent infrastructure primitives — Sep 25, 2026. Cua shipped an open-source unified computer-use driver across Linux, Windows, macOS, and Android fleets; Railway/Docker intr