Anthropic maps agent memory: five layers and a 'dreaming' pass

Its engineering post reads like a build spec for agent memory — plus ElevenLabs' new default SFX model and read-only bank data inside Claude.

Nowline NOV 7 11:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Agent memory, defined in five layers

    Anthropic's new engineering post splits agent memory into five parts — transcripts, notes, structured 'representations' like wiki pages, injected context, and an offline 'dreaming' pass that replays old sessions to fix wrong notes. The copyable wins: keep the injected layer tiny (Claude Code loads only the first 200 lines / 25KB of MEMORY.md), give each document its own page instead of one shared table (100% of facts kept vs 72% when agents shared rows), and let strong models judge what's worth saving rather than hard-coding 'record only failures.' On a 420M-token corpus, wiki-plus-search scored 61% vs 53% for search alone while saving ~20K tokens per query.

  • ElevenLabs makes SFX v3 the default — and starts a v2 clock

    SFX v3 (eleven_text_to_sound_v3) is now live for every workspace and the default in Sound Effects, Studio and Flows: it follows multi-step prompts, handles music and vocals, and renders seamless loops when you set loop: true. The catch if you build on the API — eleven_text_to_sound_v2 is deprecated and gets removed March 1, 2027, so swap your model_id before then. Pro+ asset uploads also jump from 1GB to 3GB, and audiobook projects can now start from Word .docx files.

  • Claude can now read your bank and card accounts

    A US beta for Pro, Max and Ultra links bank and credit-card accounts through Plaid, giving Claude read-only access to balances and up to 24 months of transactions to answer spending questions — no payments, transfers or investing, and Anthropic says it won't train on the data. Heads-up if you connect it: linked banks stay connected even after a paid plan lapses, until you manually unlink them.