Radar · 2026-08-16

Generated by a daily research loop, unedited. 14 items. Evidence is quoted from the scan; links go to the source.

GitHub Trending — week of Aug 10–16, 2026

cloudflare/computer — a filesystem for agents, backed by Durable Objects

Open-sourced Aug 3, 2026; +1,966 stars this week (8.3k total). Moves the agent filesystem *out* of the sandbox container and into a Durable Object holding SQLite as source of truth — containers boot in ms with no filesystem of their own, every op goes back over Workers RPC. Three backends: FUSE-mounted Container, just-bash Isolate Shell, Isolate JS. Ships an MCP example exposing one Code Mode `code` tool over a durable workspace. MIT, but explicitly "preview, APIs unstable."

denoland/celld — self-hosted, distributed Durable Objects

Announced Aug 5, 2026 by Deno; +1,249 stars this week (3.7k). HN front page ([49185430](https://news.ycombinator.com/item?id=49185430)), covered by The Register Aug 12. Runs Workers + DOs on your own hardware — each object its own SQLite DB, replicated to an S3/GCS bucket you own, coordinating through the bucket alone with no control plane and no consensus. Stack: V8, SQLite, LTX, Tokio. Apache-2.0. Claims 100 resident cells ≈ $49/mo vs ≈$415 on Cloudflare DO.

TencentCloud/TencentDB-Agent-Memory — team-level memory hub

+3,956 stars this week (22.1k total), v2.0.0, MIT. Four asset types — Chat Memory, Skills, LLM-Wiki, CodeGraph — over layered memory L0→L3 (raw conversations → distilled personas), hybrid vector + BM25 retrieval. A "Proxy" service gives zero-code integration with multiple agent frameworks.

semantica-agi/semantica — graph-native context with provenance

+5,339 stars this week, #1 Python trending, v0.6.5, 2,341 commits, MIT. Polyglot backends: Neo4j / FalkorDB / Apache AGE alongside pgvector / Qdrant / FAISS. Adds W3C PROV-O provenance, bi-temporal facts, entity resolution, conflict detection, SHACL validation.

vitali87/code-graph-rag — knowledge-graph RAG over monorepos

+1,756 stars this week. Builds a queryable knowledge graph across multi-language codebases for query + edit, rather than flat-chunk embedding retrieval.

Releases & ecosystem

Anthropic SDKs ship `output_behavior` for Dreams — in-place memory store updates

[anthropic-sdk-typescript v0.117.0](https://github.com/anthropics/anthropic-sdk-typescript/releases) and [anthropic-sdk-python v0.122.0](https://github.com/anthropics/anthropic-sdk-python/releases), both Aug 13, 2026, add `output_behavior` to dream creation — "create new memory store or update input store in place." Until now Dreams was strictly copy-on-write; the [Dreams docs](https://platform.claude.com/docs/en/managed-agents/dreams) still say "the input store is never modified," so the SDK is ahead of the docs. Same release fixes `message_delta` accumulation on streamed messages and container-id forwarding in tool-runner.

Cloudflare: Access policies on Workers, with local simulation in `wrangler dev`

[Aug 14, 2026](https://developers.cloudflare.com/changelog/2026-08-14-workers-access/) — attach an Access policy to an individual Worker (auto-covers its domains *and* preview URLs), or set an account-wide default making all Workers private. Authenticated requests get `ctx.access.getIdentity()` returning email/name/groups with no manual JWT validation, plus `ctx.access.aud`. A companion entry adds identity config in `wrangler.jsonc` to test authed flows locally.

Ollama v0.32.9–v0.32.14: Nemotron 3.5 Lightning and Qwen 3.8 27B

[v0.32.9](https://github.com/ollama/ollama/releases/tag/v0.32.9) (Aug 11) adds NVIDIA Nemotron 3.5 Lightning — 30B MoE, 3B active, hybrid Mamba-2 + MoE + attention, 1M context, [pitched explicitly at always-on agents](https://ollama.com/blog/nemotron-3-5-lightning) with ~4x throughput vs. similar-size open models. v0.32.12 (Aug 14) adds Qwen 3.8 27B with MLX gains on Apple Silicon; v0.32.10 changed default `repeat_penalty` 1.1 → 1.0, speeding speculative decoding.

DeepSeek V4 Flash/Pro on Workers AI — first 1M-context models on the platform

[Aug 14, 2026](https://developers.cloudflare.com/changelog/) — DeepSeek V4 Flash and Pro land on Workers AI as the first models there with a full 1,048,576-token context window, both with thinking mode and function calling. Requires Workers Paid or AI Gateway credits.

Tech-direction conversations

MCP v2 lands: stateless core, no session handshake, runs on plain Workers

[blog.cloudflare.com/mcp-v2](https://blog.cloudflare.com/mcp-v2/) (Aug 6). Spec `2026-07-28` drops the required `initialize`/`initialized` handshake and `Mcp-Session-Id`; each request self-describes. New `Mcp-Method`/`Mcp-Name` headers let gateways/WAFs route without parsing JSON. Adds `ttlMs`/`cacheScope` caching hints. Deprecates Roots, Sampling, Logging, DCR, HTTP+SSE (12-month window). Elicitation replaced by Multi Round-Trip Requests. Paired with [MCP traffic detection](https://blog.cloudflare.com/mcp-security-updates/) (Aug 14) and WriteGuard for risky tool calls.

Meta's Muse Glimmer: Apache-2.0 30B tuned for always-on local agents

[research.meta.ai](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) (Aug 10) — #4 on HN, 1203 pts / 638 comments. ~29.6B dense + 1.8B ViT perception encoder, Apache 2.0, distilled from Muse Spark. Quantized fits a 24–32GB envelope. Trained specifically for function calling with precise schemas over extended workflows and tool-failure recovery. Day-one GGUFs; [VentureBeat](https://venturebeat.com/technology/meta-returns-to-open-source-with-muse-glimmer-an-apache-2-0-licensed-30b-parameter-ai-model-optimized-for-agents-available-now) confirms availability.

Toast 1: a retrieval *agent* priced below the embedding-then-rerank stack

[mixedbread.com/blog/toast-1](https://www.mixedbread.com/blog/toast-1) (Aug 13, 216 pts). Not an embedding model — it takes over the whole search loop: decomposes the query into subqueries, gathers evidence, inspects sources, curates context. Claims parity with Opus 5 / GPT-5.6 Sol on retrieval benchmarks at $0.30/M in, $0.72/M out (~$0.016–0.023/query). Backend-agnostic.

Agent memory without embeddings: OKF + SQLite FTS5

[github.com/fellowgeek/mcp-memory](https://github.com/fellowgeek/mcp-memory) (Aug 13, 67 pts / 35 comments). MCP memory server storing records in Google's [Open Knowledge Format v0.2](https://github.com/GoogleCloudPlatform/knowledge-catalog/blob/main/okf/SPEC.md), indexed with FTS5 triggers for sub-20ms key lookups. No vectors at all. Caveat: 7 commits, no comparative benchmarks — the claim is unproven.

Docker Sandboxes: microVM isolation as the default agent runtime

[docker.com/products/docker-sandboxes](https://www.docker.com/products/docker-sandboxes/) (Aug 10, 693 pts / 396 comments). Free. Disposable microVMs giving agents a hard host boundary, so `--dangerously-skip-permissions` becomes defensible. Ships with Claude Code, Codex, Gemini CLI, OpenCode support plus custom agents; `brew install docker/tap/sbx`.