The memory thesis ships as a product

Tencent’s Hy-Memory-30B-A3B is a specialist writer for external memory frameworks — competitive with Claude Opus 5 on the real post-phantom instruments; DeepSeek’s -1017 refresh posts the largest agent-suite jump we’ve ledgered at unchanged prices; and Sakana’s Agentic Scientist goes public with its own orchestration ablation.

Hy-Memory-30B-A3B (Apache-2.0, Oct 15) turns long-term-memory writing into a specialist 30B-A3B model scored on LoCoMo/LongMemEval — the instruments that replaced the phantom benchmark; V4.1-Flash-1017 lifts Terminal-Bench 4.0 from 31.2 to 42.3 with pricing unchanged and opens DeepSeek’s in-house Search API the same day; Real-SWE 1.1 crowns the cheap model (GPT-6.1 Sol at $0.79/rollout); Sakana’s Agentic Scientist ships public and runnable, ablation showing the orchestration layer — not the base model — carries scientific autonomy. Issue 14, coverage window 15–18 October 2026.

Week of October 15–18, 2026 · Issue 14

The Week in One Sentence

The memory-systems thread this publication has tracked since Issue 1 shipped as a product this week — Tencent’s Apache-licensed specialist whose entire job is deciding what an agent remembers, competitive with the frontier on exactly the instruments that replaced our phantom — while DeepSeek’s dated refresh posted the largest same-endpoint jump in our ledger without touching prices, Real-SWE’s first scoring of the launch generation crowned its cheapest model, India shipped the missing evaluation layer as one ten-benchmark drop, Sakana’s Agentic Scientist went public carrying its own orchestration-beats-model ablation, and the ExtractBench full read quietly widened the extraction collapse from a China story to an everyone story.

---

The Big Story: Hy-Memory-30B-A3B — memory writing becomes a specialization

What happened. (China.) Tencent’s Hunyuan lab released Hy-Memory-30B-A3B (created October 15; FP8 variant days later) — a 30B-total / 3B-active MoE whose entire job is the write path of external memory: deciding what to retain from a conversation or task, how to represent it, and how to update the persistent bank, across Chat, Working, and Coding scenarios. It drops into existing frameworks — Mem0, MemOS, Hy-Memory — and transfers to a held-out harness (EverMemOS). Apache-2.0 (license field read verbatim). Training is two-stage: Memory On-Policy Distillation, then what the paper calls Attributed Reinforcement Learning for Long-Term Memory Writing (arXiv:2610.22196 — full read queued; the code ships at github.com/Tencent-Hunyuan/Hy-Memory).

What the evidence says (vendor-run, on real instruments — the detail that matters most). The benchmark table scores writers inside four harnesses using LoCoMo, LongMemEval, PersonaMem-v2, and SPS/code benches — the exact post-phantom instruments this desk named two weeks ago after MarathonMemBench evaporated. Macro average: Hy-Memory 55.49, ahead of Claude Opus 5 at 54.62 — a 0.87-point margin on a fixed 20% subsample, which we label precisely: competitive with the frontier, not proven above it, and every number is vendor-run. The table’s honest spine is the ablation-shaped rows: +7.8 points separate the family’s base row (Hy3-A3B-Base, 47.66) from the trained writer in the same table — specialization gain on a small MoE anyone can host at ~60 GB.

Why this matters. This desk has tracked the memory-systems thesis through five architecture iterations (Engrams, indexers, KV compression) — all inside models. Hy-Memory is the thesis crossing the boundary: memory as an external, swappable, specialist component, with the general model demoted to a consumer of the memory bank. It is the SAS/EVIE/WeVisDoc pattern (small frozen-base specialists) applied to the agent stack, and it arrives the same month the benchmark instruments finally matured. The open questions are exactly the ones our ledger holds: whether “attributed” reinforcement means provenance-carrying writes (the grounding-shaped hole ExtractBench just exposed), and whether the 0.87 margin survives independent writers’ evals.

What I would test. Swap your framework’s writer (most people use the general chat model) for Hy-Memory under Apache-2.0, on a fixed conversation corpus, and measure recall and pollution: correct retrievals per hundred memories written, and stale-wrong memories that should have been updated but weren’t. The write path is where memory systems fail quietly; nobody’s benchmark column shows pollution yet — be the first table.

---

Elsewhere in AI Engineering

V4.1-Flash-1017: the dated refresh, the Pareto claim, and a search API (China) — Worth Testing

What happened. (Labelled update on the Issues 4/6 lineage.) DeepSeek’s October 17 changelog upgraded deepseek-flash to V4.1-Flash-1017 — “a minor update,” pricing and calling method unchanged (read day-of from the primary; HIGH-cost, verified). Vendor deltas are anything but minor on the agent suite: Terminal-Bench 4.0 from 31.2 to 42.3 (+11.1 — the largest jump this ledger has recorded on a same-endpoint refresh, landing exactly on the science gap the V4.1 report conceded), DeepSWE v1.1 74.2→77.0, ProgramBench 20.3→29.1, with shorter, faster outputs — the card’s phrase is “advancing the performance–cost Pareto frontier,” and at unchanged prices that frontier moved for free. The release also drops the Jinja chat template (resolving last issue’s card contradiction in the opposite direction) in favor of a Python reference encoder plus a Rust/Python “deepseek-recipe” toolkit. Same day, the same changelog opened DeepSeek’s in-house Search & Fetch API to the public — the search engine behind their stacks, now a product.

Why it matters. The dated-refresh convention (-0731, -0813, -1017) is becoming DeepSeek’s release rhythm: silent-ish, same endpoint, same price, better model — which makes changelog-watching a cost-of-goods strategy, not news consumption. The search API matters more than it looks: agent stacks pay for search as a line item, and a lab-priced entrant with its own index changes that budget. Independent evals of -1017 don’t exist yet; the +11.1 is vendor-run.

DSpine: speculative drafting through depth (China) — Worth Reading

What happened. (Dating disclosed: paper September, arXiv:2609.36173 — the inference-code release landed in-window, October.) Tencent’s DSpine is a 1B parallel drafter for Qwen3-8B under Apache-2.0: instead of drafting tokens sequentially, all positions draft in parallel while causality “unfolds over network depth” — each draft layer injects the predecessor’s predicted feature. Acceptance length averages 7.83 at temperature 0 versus DFlash2 (7.55), DSpark (7.44), Domino (7.05), DFlash (5.92) — with the baselines retrained by the authors for apples-to-apples, a discipline worth calling out because this field rarely bothers. SGLang recipe ships; training code and a 27B drafter are on the roadmap.

Why it matters. The draft-model line keeps compounding quietly (DSpark inside V4.1, Draft-KV’s inter-model bridge in Issue 9, now DSpine’s depth-parallel twist): acceptance-length points are throughput you don’t pay for in quality, and the retrained-baseline table is the format every spec-decoding claim should arrive in.

The Indic eval battery arrived in one drop (India) — Worth Reading

What happened. (India.) October 16, per the org’s activity feed: ten evaluation datasets published in a single day — indic-mmlu-pro, indic-aime-2026, indic-swe-bench, indic-hle, indic-math-500, indic-gsm8k, indic-mbpp, indic-humaneval, indic-jee-bench, indic-arc-challenge. We read the largest card in full: indic-mmlu-pro is MMLU-Pro translated into 14 Indic languages (Assamese through Urdu, plus English; 10-option MCQA with chain-of-thought rationales preserved and per-language columns), MIT-licensed (license field read verbatim), 10K–100K rows in parquet.

Why this matters. This is the missing layer: evaluating models in Indian languages on hard tasks — reasoning, competition math, code, SWE, HLE, even the IIT entrance exam — rather than translating English results after the fact. Ten benchmarks landing as one battery is infrastructure, not gardening, and it lands the same month this publication has been documenting how thin the independent layer is everywhere else. The caveats sit on the card, or rather don’t: the translation pipeline and any human verification are undocumented on the card we read — a translated benchmark inherits translation quality as noise — and no accompanying paper is visible yet. Only indic-mmlu-pro’s license was verified; we claim nothing suite-wide. Those gaps are the distance between an eval battery and a trustworthy one; the methodology note is the thing to watch for.

Go deeper: indic-mmlu-pro · org listing

Real-SWE 1.1 finally scored the launch generation — and crowned the cheap one (US) — Worth Reading

What happened. (Labelled update on this desk’s longest-running instrument watch.) Real-SWE shipped version 1.1 (page-dated October): 25 licensed private-codebase tasks across five languages, 16 models, five runs per task with 95% intervals and per-rollout costs — and the new entries are the entire launch generation this publication has tracked since September: Gemini 4 Argon, GPT-6.1 Sol, Grok 4.7, Opus 5.5, Muse Wonder. The result: GPT-6.1 Sol takes #1 at 47.20% — at $0.79 per rollout — ahead of Fable 5.1 (45.60% at $14.24), GPT-6 Astra (44.80% at $3.78), and the tied Opus 5.5 and Muse Wonder 1.0 (42.40%). Gemini 4 Argon places 6th at 39.20% ($4.66); Grok 4.7 manages 25.60% and GLM-5.3 26.40% — but those two are cross-version readings (1.1 is 25 tasks and five runs, not 1.0’s ten and eight), and we print no 1.0→1.1 delta as comparable. The value row belongs to a China-lab model: DeepSeek V4.1 Flash, 7th at 38.40% for $0.19 a rollout.

Technical signal. The cost column is now the story: an 18× spread separates Fable 5.1 from the leader for 1.6 points of resolution, and GPT-6 Luna scores 10.40% at $0.08. The failure taxonomy is the diagnostic gift: “unverified assumption” dominates at 33.4% of failed runs, half of Argon’s failures are “misunderstood task” — a comprehension-shaped signature, not an execution one — while the OpenAI pair skews 53–55% unverified-assumption. No Sonnet or Haiku 5.5 entries exist; the Claude family appears as Opus 5.5 and Fable 5.1 only.

Why it matters. The top-of-instrument disagreement is now three cycles old — AA crowns Opus 5.5, Real-SWE crowns GPT-6.1 Sol — but 1.1 adds the axis AA lacks: what the win costs. On the one instrument with a per-rollout price column, Issue 13’s “near-Astra at a fifth of the price” resolves to above Astra at a fifth of the cost, and Argon’s vendor-SOTA framing meets a 39.20% placement with a task-comprehension failure signature. Caveats labeled, as always: author-run throughout, five runs not eight, highest-effort setting, still no paper.

Go deeper: Real-SWE 1.1

Sakana’s Agentic Scientist went public — runnable, and ablated against itself (Japan) — Worth Reading

What happened. (Japan. Dating: the paper is October 13 — pre-window, disclosed; the code went public October 15, this window’s event.) Sakana released AI-Scientist-v3 — The Agentic Scientist (arXiv:2610.18933, 88 pages, Robert Tjarko Lange) — a distributed pipeline that “treats coding agents as AI scientists”: from a task.md, agents propose ideas, run sandboxed experiments against a rubric, iterate, then the pipeline runs baselines and ablations, a sealed final evaluation, post-hoc analysis, and writes the paper — six phases under one run directory, usable as a co-scientist tool or fully autonomously. Apache-2.0, and 342 stars within three days — the fastest uptake of any Japan artifact this publication has ledgered. Its sibling went public the same morning: AutoIllustrate (Apache-2.0, 93★) turns a methodology brief into an editable .drawio vector figure with an optional image-model restyle pass — “the generative model only touches the final restyle” — and is built into the Agentic Scientist’s Docker image; it drew v3’s own diagrams.

Technical signal. The claims that matter come in pairs. First, the venue claim: in co-scientist mode, a development version of the system helped produce a paper accepted for publication at ALIFE 2026 — vendor-stated, we have not located the paper independently. Second, the ablation: reliable scientific autonomy “relies not just on base model capability but also on the cross-agent orchestration layer” — the harness-beats-model result arriving as a vendor’s own controlled finding, in the same issue whose Real-SWE table crowns a harness-native entry and whose memory Big Story is a +7.8 specialization gain. And the README ships the season’s most candid security note: the agents launch with full tool permissions (--allow yolo), can run arbitrary commands and inherit your API keys — “keep runs inside the Docker image… only run tasks and sources you trust.” The candor pattern (MASS’s decline tables, now this) is becoming Sakana’s house style; the uptake-without-terms pattern (MASS unlicensed at 278★) is not, so far — v3 shipped Apache-2.0 from day one.

Why it matters. Automated research is the load-bearing claim of the agentic era, and this is the most runnable version of it from a major lab: agsci CLI, Docker-per-phase, GPU presets, doctor checks. For teams evaluating it, the two checks that matter are the ones the paper’s own ablation invites — does the orchestration layer survive your tasks, and does the sealed-evaluation discipline hold when the agent wrote the evaluation? Both are one-weekend experiments.

Go deeper: repo · paper · AutoIllustrate

---

Research Radar

The ExtractBench full read: the collapse is everyone’s (China desk; labelled update on Issue 13)

Read in full this cycle (the stronger basis review asked for), and the graduation claim holds — wider. The paper’s own 14-system table shows the short→long extraction collapse is cross-ecosystem: Gemini 3.5 Flash falls 87.9 → 27.9, GPT-5.4 Nano 77.4 → 35.8 — the same neighborhood as the card-reported GLM-5.3-Flash (27.83) and V4.1-Flash (22.23) longs¹. Coding agents hold (Claude Code Opus 4.8: 90.1 → 88.1); specialized APIs hold (LlamaExtract Agentic Plus never below 90). The slices sharpen it: on enormous tables, every VLM scores under 10%; failure is recall-driven — truncated record lists attributed to context limits. Ground-truth construction is genuinely strong (unanimous multi-system agreement; backwards-built synthetic lists with exact-by-construction truth; 169 human-verified forms). Two caveats at full strength: all five authors are runllama.ai-affiliated with no conflict-of-interest disclosure in the paper — failure modes credible, rankings marketing — and Contextual AI maintains an unrelated benchmark with the same name. Our Issue 8 framing (“the efficient-context generation has a failure signature”) widens to: general-purpose VLMs have the signature; the efficient-context models just sit at the bottom of the same slope.

---

Repo Radar

Tencent-Hunyuan/Hy-Memory (China)

• What’s actually in it: the writer model (~60 GB BF16, FP8 variant), framework integration examples (vLLM + OpenAI-compatible, no-think JSON extraction), training code per the two-stage recipe. Inspection depth: card + README-level; code not executed this cycle.
• What you can run: a memory-writer service tonight, under Apache-2.0, in any of the supported frameworks.
• What’s missing: the paper’s full method read (queued); independent writer evals; a pollution metric.
• Why open it: the first memory-path specialist we’ve ledgered from a major lab — the component your agent stack didn’t know it could swap.

---

My Read on the Week

(Synthesis across the four desks that ran — marked as interpretation throughout; every figure traces to a ledger row.)

1. The stack is decomposing into specialists — memory is just the newest organ. Parsing, retrieval, drafting, now memory-writing: each got a small, permissive, focused model this quarter, each beating or matching a frontier generalist on its slice. Japan’s items cut the same direction twice — AutoIllustrate shrinks figure-drawing to a brief-to-vector specialist, and the Agentic Scientist’s own ablation says the orchestration layer, not the base model, carries autonomy. When Hy-Memory’s own table shows +7.8 points from training the skill rather than the model, the architecture question for agent builders stops being “which LLM” and becomes “which organs.”

2. Free-lunch updates are a strategy, not an event. V4.1-Flash-1017’s unchanged prices with +11.1 on the hardest agent benchmark is the third such quiet refresh this quarter. The buyer’s move is infrastructure: pin to dated endpoints, watch changelogs, and re-benchmark on refresh cadence — because your incumbent’s moat (integration cost) is exactly what these updates attack at zero price delta.

3. Instruments finally grew teeth — in both directions. Real memory benchmarks (LoCoMo, LongMemEval) now anchor vendor claims the week after a phantom one evaporated; ExtractBench’s full table both corroborates the collapse pattern and exemplifies vendor-adjacent evaluation (authors’ product on top, no disclosure); Real-SWE 1.1 adds the axis every other table lacks — what the win costs — and India’s ten-benchmark battery builds an evaluation layer where none existed, translation caveats and all. The reader’s toolkit is now stable: check the instrument’s provenance before its rankings, every time.

---

What I’d Watch Next

1. Independent memory-writer evals — the 0.87 Opus-5 margin is precisely the size that independent runs flip; first third-party table settles it.
2. -1017 under independent instruments — TB 4.0 +11.1 vendor-run is extraordinary; the first independent agent-suite run either confirms the science-gap closure or prices the marketing.
3. First independent runs of the Agentic Scientist — the orchestration-beats-model ablation is vendor-run and rhymes with the two watches above; the first third-party reproduction — and the independent location of the ALIFE 2026 paper — tests both the claim and the candor pattern.

---

One Thing I’d Experiment With This Week

Instrument your agent’s memory pollution rate — before and after a specialist writer. Hy-Memory’s table scores recall-shaped metrics; nobody scores the failure mode that actually poisons agent stacks: stale or wrong memories that keep firing.

• Setup: your existing agent + memory framework, a fixed two-week replay of real conversations. Arm one: the general model as writer (your current default). Arm two: Hy-Memory-30B-A3B as writer, same framework, same read path.
• Conditions: after the replay, run a 100-query evaluation set against the memory bank — queries with known ground truth including outdated facts (the user changed their mind mid-corpus) and never-true near-misses.
• Measure: recall on current facts; pollution rate (fraction of retrieved memories that are stale or wrong — a metric none of these vendor tables carry); bank size and cost per thousand conversations written. Decision output: whether the specialist writer earns its 60 GB on correctness, not just recall — and the first pollution-rate number in the wild, which is worth publishing regardless of which arm wins.

---

Corrections

[Japan] 2026-10-18 — Issue 9’s PLaMo deployment guidance omitted a term that gates every commercial use. Issue 9’s license item told readers that below the ¥1B revenue ceiling (~US$7M) “you are licensed; above it, the contact form.” The rendered license page (read October 14, recovered from the Japan desk run interrupted that day) confirms the ceiling, naming, and output-inheritance terms Issue 9 quoted — and adds what the earlier static-text extraction of the JS-rendered page missed: Article 3.3 requires prior registration on PFN’s designated form for any commercial use, on top of the revenue condition; Article 1.9 defines commercial purpose broadly enough to include hosting services, APIs, and use inside your own organization; and Article 8.1 caps PFN’s aggregate liability at ¥500 absent intent or gross negligence. The published guidance was therefore incomplete in a decision-relevant way: an under-ceiling commercial deployment is licensed only after registering, not automatically. Root cause: guidance shipped from an unrendered extraction of a page that requires JavaScript — the same render-or-don’t-cite class this playbook’s instrument checks document. Re-verified against the PLaMo Community License (rendered), read October 14; ledgered as JP-8–JP-10 in research/2026-10-18/claim-ledger-jp.md.

---

All vendor-reported figures are labeled as such; if independent evidence doesn’t exist yet, that’s stated rather than papered over. This issue’s desk ledgers live under research/2026-10-18/.

Footnotes

1. ExtractBench’s comparator set (Gemini 3.5 Flash, GPT-5.4 Nano, Codex GPT-5.5, Claude Code Opus 4.8) and Hy-Memory’s (Claude Opus 5) are different generations from this issue’s other tables and from Issues 1–13’s spreads — cross-table score comparisons across comparator vintages remain not apples-to-apples (EDITORIAL.md §6 discipline).