Cognee v1 Memory, GPT-5.6 Sol Preview, UI-to-Code Workflows — AI Daily Jun 26

468 messages · 76 active members

468

messages

active members

@jcartu, @Wootbro, @samb69

top contributors

Overview

The biggest drop of the day was Cognee v1 from @Max_Bernstein, a memory layer claiming 145% better long-context recall than Opus 4.8 and GPT 5.5, an effective 100B-token context window, and 6.9x lower cost — pitched as a bolt-on for existing agents rather than a new platform. It landed in a community already declaring 'memory frameworks are the new JS frameworks' (@Anonymoushat) and rethinking model dependence as @jcartu NVFP4-quants a new Orinth model (Qwen 3.5 397b + vision stack) on his RTX 6000, arguing that if its intelligence claims hold, it's a major leap given current parity needs 768B+ params. OpenAI previewed GPT-5.6 Sol, splitting the room between fans of the personality-driven naming and builders who'd rather see launches than teases. On the practical side, @tvrxm kicked off a long thread on getting Codex to match designed UI screens 1:1 — consensus was to build a real design system (components, typography, spacing) rather than feeding screenshots, with Gemini AI Studio, Claude Design, and Google Stitch recommended. @realcrischico's question about sharing skills between Codex and Claude produced a clean answer: symlink ./.claude to ./.codex and use `@AGENTS.md` includes in CLAUDE.md. Hermes operators kept debugging compaction: @samb69's /goal stopped at 20 turns despite a 600-turn cap until @samtome's compact-timeout fix worked 'surprisingly well,' and @jcartu warned that 1M-context users must cron hindsight memory dedup 3x daily or context poisoning silently breaks runs. Infra threads covered Cerebras hitting 750 tok/sec (with KV-cache memory caveats), Seedance video gen pricing (Kie ~$1.40/15s vs BytePlus ~$1.10/video at $55 per 13M tokens), and a geopolitics tangent on Elon's AI race play via Grok, SpaceX chips, and orbital data centers.

Topics

Cognee v1: 100B-Token Memory Layer for Agents9 msgs

@Max_Bernstein shared Cognee v1, claiming 145% better long-context recall than Opus 4.8/GPT 5.5, a 100B-token effective context window, and 6.9x lower cost. It's positioned as a drop-in memory layer for existing agents, landing amid wider community talk that 'memory frameworks are the new JS frameworks' and ongoing frustration with context poisoning.

OpenAI Previews GPT-5.6 Sol12 msgs

OpenAI teased GPT-5.6 Sol, drawing split reactions: some builders liked the personality-driven naming over anthropomorphized 'ask Claude' patterns, others wanted shipping over teasing. Skepticism centered on whether Sol will meaningfully beat competitors on benchmarks like terminal-bench.

UI-to-Code Workflows for Pixel-Perfect Screens20 msgs

@tvrxm sparked a long discussion on getting Codex to match designed UI screens 1:1 after burning tokens for ~5/10 fidelity. Recommendations included Gemini AI Studio, Claude Design, Google Stitch, and a lazycodex/stitch workflow. Consensus: build a design system with components, typography, and spacing rules instead of feeding raw images to the agent.

Hermes /goal Loops, Compaction & Memory Hygiene7 msgs

@samb69 hit the classic Hermes problem of /goal stopping at 20 turns despite a 600-turn cap; @samtome's fix was to extend the compact timeout, which worked 'surprisingly well.' @jcartu added that 1M-context users must cron their hindsight memory stack 3x/day for dedup and consolidation, since context poisoning silently broke his own runs before automation.

Codex/Claude Skill Sharing & Local Quants on RTX 600014 msgs

@realcrischico asked how to share skills, CLAUDE.md, and config across Codex and Claude — the answer: symlink ./.claude to ./.codex and use `@AGENTS.md` includes so each harness layers on a shared base. Meanwhile @jcartu is NVFP4-quanting a new Orinth model (Qwen 3.5 397b + vision stack) on his RTX 6000, noting Cerebras hits 750 tok/sec but ASIC memory limits constrain KV cache and context length.

Key Takeaways

Cognee v1 claims 100B-token context and 145% better recall than Opus 4.8/GPT 5.5 as a drop-in memory layer — worth testing on existing agent stacks.
Fix Hermes /goal stopping early by extending the compact timeout; the 600-turn cap is meaningless if compaction trips first. Cron hindsight memory dedup 3x daily on 1M-context runs.
For pixel-perfect UI generation, build a real design system (components, typography, spacing) before feeding screens to Codex — raw screenshots top out around 5/10 fidelity.
Symlink ./.claude to ./.codex and use `@AGENTS.md` includes in CLAUDE.md to share skills across both harnesses with harness-specific overrides.
Seedance via BytePlus ($55 / 13M tokens ≈ 50 videos at 480p) is meaningfully cheaper than Kie's ~$1.40 per 15s clip for high-volume ad creative; Cerebras hits 750 tok/sec but validate KV-cache limits first.