+4,304 stars this week (9,304 total), MIT, commits as recent as Aug 19. Hit #1 on GitHub Trending Aug 10. Ships an MCP server (`python -m semantica.mcp_server`) with entity-extraction, decision-recording and graph-query tools, and backs onto **Neo4j + pgvector** directly. Pitch is deterministic graph construction — no LLM in the provenance path.
Team Memory announced **Aug 13, 2026**; +3,207 stars this week (23.1k total, 20k in 90 days), MIT. Four memory asset types — Chat Memory (L0–L3 distillation), Skills, LLM-Wiki, Code-Graph — with a **governance layer** so a teammate's agent reads what yours learned, minus anything marked private. Installs via proxy, no code change to Claude Code/Codex.
+1,611 stars this week off a 1,886 total — near-vertical, brand-new. Apache 2.0, NVIDIA copyright. Routes across providers, converts between Anthropic Messages / OpenAI Chat / Responses formats, works with Ollama and vLLM endpoints, emits operational metrics. Explicitly labeled pre-alpha, not for production.
+1,148 stars this week (3,004 total), MIT, Rust. Captures agent lifecycle hooks (prompts, tool calls, session boundaries) and compiles them into a **git-versioned markdown wiki** indexed by SQLite FTS5. Runs as MCP + HTTP server; support matrix covers Claude Code, Codex, Cursor, Gemini CLI. Pitch: quit Claude Code mid-task, resume in Codex without re-explaining.
+985 stars this week (29.8k total). Stores memory/resources/skills under a `viking://` virtual filesystem; agents retrieve with `ls`/`tree`/`find` rather than vector search, over L0/L1/L2 abstraction tiers loaded on demand. MCP client integration. **AGPLv3** core (Apache 2.0 only for CLI/examples).
anthropic-sdk-typescript v0.117.0 and anthropic-sdk-python v0.122.0 (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 fixed `message_delta` accumulation in streaming and tool-runner container-ID forwarding.
Ollama v0.32.7/v0.32.8 (Aug 10) shipped Muse Glimmer on Apple Silicon MLX then all platforms. 30B MoE distilled from Muse Spark, 128K context, 1.8B perception encoder for native image input, four reasoning-effort levels, Apache 2.0. Ollama's own post says it can power Claude Code–style always-on local agents.
Released Aug 13, 2026; ~110 bug fixes. Headline: a self-deadlock while replaying WAL generated by an *older minor version* — a regression introduced in the previous minor set that can leave a standby stuck. Also fixes parallel GIN index builds reporting uninitialized `reltuples`, and tzdata 2026c.
Aug 18, 2026. New `container` field on exports links a container to a DO bidirectionally, making `containers[].class_name` optional. Separately, Wrangler now defaults the compatibility date to its bundled workerd's release date instead of today's date. Plus `.env` casing fix on Windows and D1 CRLF/lowercase-`end` SQL parsing fixes.
Aug 14: Access can now protect an individual Worker across all its domains, or every Worker account-wide by default with exemptions; `ctx.access.getIdentity()` works in `wrangler dev`, and policies are manageable via the Workers API. Aug 14: DeepSeek V4 Flash/Pro on Workers AI — first models there with a full 1,048,576-token context. Aug 17: Qwen 3.8 27B (vision, 262K context, function calling).
[HN #1 Aug 14, 1430 pts / 791 comments](https://news.ycombinator.com/item?id=49299605) on the [HF card](https://huggingface.co/Qwen/Qwen3.8-27B-FP8); Simon Willison's [follow-up](https://simonwillison.net/2026/Aug/16/qwen-38-27b/) (788 pts) says it's "excellent, but defaults to wildly overthinking things." 262k native context, Terminal-Bench 2.1 73.0 (up from 63.4), SWE-bench Pro 61.7, beats Opus 4.6 on 15/19 overlapping tests, runs quantized on 24GB VRAM.
[Anthropic research, Aug 13](https://www.anthropic.com/research/multiagent-systems); [HN 196 pts / 138 comments](https://news.ycombinator.com/item?id=49316271). Names five failure modes: poor coordination (agents abandon conflicting PRs or self-isolate), conformity risk ("when one agent makes a bad decision, many make the same one"), collusion, epistemic brittleness, and escalation under incompatible goals. Shared forums/repos did *not* fix coordination.
[HN Aug 14, 983 pts / 868 comments](https://news.ycombinator.com/item?id=49296740). Recurring complaints: verbosity that obscures meaning, over-commented code, CLAUDE.md instructions ignored, benchmark-shaped regressions vs 4.6/4.8, cancelled $200/mo subs. Not unanimous, but unusually dense with practitioners citing specific harness behavior.
[deepseek.com/harness, Aug 13](https://deepseek.com/harness/en/); [HN 739 pts / 309 comments](https://news.ycombinator.com/item?id=49285244). Cordis kernel mounts models, tools, skills, sessions, sandboxes, storage, loops as swappable plugins; append-only session log with fork/resume/replay. MIT. No MCP support mentioned on the page.
[HN Aug 17, 567 pts / 416 comments](https://news.ycombinator.com/item?id=49334209); coverage via [SiliconANGLE](https://siliconangle.com/2026/08/17/cursor-origin-code-hosting-service-to-compete-with-github/). Beta to paid users, claims 296k clones and 22 commits/sec per repo, GitHub sync. Landed during a seven-hour GitHub outage.