AlphaSignal — OptMem's compressed agent memory, Kimi K3's 2.8T open weights, SSI's $5B NVIDIA deal (Jul 28 2026)

Why this is in the vault

Third entry in an ongoing vault watch on agent-memory tooling (following [[2026-05-18-alphasignal-agentmemory-92-percent-fewer-tokens]] and the Qwen sleep-time compression item in [[2026-05-27-alphasignal-qwen-1m-context-sleep-memory-compression]]). OptMem is a small, concrete open-source instance of exactly the context-rot mitigation pattern CLAUDE.md hard rule #4 already codifies for this agent, plus two other items (Kimi K3's open-weight release, SSI's NVIDIA compute deal) worth a data point each for RDCO's open-model and compute-economics tracking.

Mapping against Ray Data Co

OptMem is the same problem CLAUDE.md hard rule #4 solves, attacked from the opposite side. Rule #4 keeps this session's context bounded by routing long artifacts (>5KB) through subagents so raw text never enters parent context — exactly what this note-writing task itself just did with the 91KB Gmail thread JSON. OptMem instead keeps memory bounded across sessions: it appends everything to an append-only LOG.txt (never deletes, so nothing is silently lost — a stronger guarantee than lossy /compact) but builds a tree of compressed summaries so an agent always reads a fixed ~208-line document regardless of how much history has accumulated. That's a middle path between the two techniques ~/.claude/state/working-context.md and Thariq's /compact vs /clear distinction ([[2026-04-15-thariq-claude-code-session-management-1m-context]]) already represent for Ray: /clear requires a human-written brief, /compact is one-shot lossy summarization, and OptMem's tree-of-summaries is closer to automated, non-destructive compaction. Same open question flagged in the prior agentmemory note applies here too — automated compression may not preserve what a human engineer would judge load-bearing, so it's a sniff-test candidate, not an adoption decision, until paper-traded against working-context.md on a fixed task.

Kimi K3 (2.8T params, open-weight, 1M context, 88.3 Terminal-Bench) extends the open-weight-frontier-model watch already running in the vault (prior Kimi/Moonshot entries: [[2026-07-20-alphasignal-depth-anything-kimi-k3-open-weights]], [[2026-07-17-alphasignal-moonshot-delta-attention-video-physics]]). The license detail worth flagging for any future client-facing deployment advice: free for personal/small-company use, but companies earning >$20M/yr need a commercial license, and 100M+-user platforms must display Kimi K3 branding — a real constraint if RDCO ever recommends self-hosting Kimi for a client past that revenue line.

SSI's ~$5B NVIDIA investment (Bloomberg-sourced, unlocking Vera Rubin GPU access atop SSI's existing Google TPU stack) is a macro compute-economics data point, not a direct tooling hit — logging for the investing/compute-cycle watch rather than the harness-thesis line.

Curation section

• Top news — SSI/NVIDIA: NVIDIA invests ~$5B (Bloomberg estimate) in Safe Superintelligence (Ilya Sutskever), giving SSI access to NVIDIA's next-gen Vera Rubin GPU platform for a claimed 10x compute increase within 12 months, layered on top of SSI's existing Google Cloud TPU infrastructure.
• Sponsored — WorkOS: MCP server giving AI agents dashboard-level access to SSO/auth config via OAuth with scoped tokens, positioned against handing agents a master API key.
• Top model — Kimi K3: Moonshot AI, 2.8T parameters, open-weight, natively multimodal (text/image/video), 1M-token context, scores 88.3 on Terminal-Bench (competitive with Claude/GPT-5 class), runs on vLLM/SGLang/Ollama/Docker. Non-commercial-leaning license: free under $20M/yr revenue; 100M+-user platforms must display branding.
• Sponsored — Anyscale: Ray Summit + vLLM Conference promo, discount code, standard paid CTA.
• Top repo — OptMem: plug-and-play agent memory system, two text files + six shell commands (memo wake, memo note, memo sleep, memo recall, memo forget). All memory appended to LOG.txt (never deleted); a tree of compressed summaries keeps the agent's working read to a fixed ~208 lines regardless of history length. Works with Claude Code, Codex, or plain shell; 3-line install.
• Signals (6 one-liners): (1) Kimi open-sources FlashKDA attention kernel, cuts prefill time up to 2.2x; (2) Augment Code "Agentic SDLC" — sponsored; (3) an unnamed open-source MCP server claims a 95% cut in AI coding token costs across 48,000 developers — no tool name given in the issue body, tracking-link-only, not independently verified here; (4) a cited LangGraph paper argues graph engineering should replace ad hoc prompt control flow; (5) a new 3B model reportedly beats Qwen-9B and Gemma-12B on agentic-task benchmarks; (6) Kimi open-sources the distributed agent-training system behind K3.

Deep-fetch note: zero deep-fetches performed. Every link in this issue (including the OptMem and MCP-token-reduction items) routes through AlphaSignal's own tracking redirector (app.alphasignal.ai/c?...) rather than a bare third-party URL, and the captured HTML's query-string encoding was corrupted in transit (equals signs replaced with control characters), making the redirect links non-functional as fetched. No specific repo/org name was recoverable for the 95%-token-reduction MCP server beyond the blurb text. Per the skill's curation-mode link rules, this is a legitimate zero — nothing here cleared the third-party-plus-specific-hook-plus-working-URL bar.

⚠️ Sponsorship

Three disclosed third-party paid placements this issue, standard AlphaSignal rotating-ad-pool pattern (same pattern documented for CFO Secrets and Mostly Metrics in the process-newsletter README): WorkOS (MCP-based auth-agent server pitch), Anyscale (Ray Summit / vLLM Conference event promo), Augment Code (Agentic SDLC pitch). None of the three sponsors are the source of, or tied to, the featured OptMem/Kimi K3/SSI items — those read as independent third-party reporting. No AlphaSignal self-promotion (no AlphaSignal Pro/course/job-board plug) in this issue.

Related

• [[2026-05-18-alphasignal-agentmemory-92-percent-fewer-tokens]] — prior AlphaSignal agent-memory item; same "context-rot inverse problem" framing, same open question about heuristic vs human-judged compression
• [[2026-05-27-alphasignal-qwen-1m-context-sleep-memory-compression]] — earlier sleep-time memory-compression pattern in the same watch line
• [[2026-04-15-thariq-claude-code-session-management-1m-context]] — first-party source for CLAUDE.md hard rule #4 and the /compact vs /clear distinction OptMem sits between
• [[2026-07-20-alphasignal-depth-anything-kimi-k3-open-weights]] — prior Kimi K3 coverage in the vault's open-weight-model watch
• [[../08-tooling/rdco-state-ownership-architecture]] — RDCO's current state/memory architecture that any adopted memory substrate would sit inside