Test what the
score hides.

AtMem reports that a fine-tuned memory-write classifier improved its offline score while failing the practical job of retaining useful information. Its production selection process blocked deployment.

V2 reached 73.79% fully correct transitions on a sealed evaluation, but an always-decline-storage baseline scored 91.10%. In a matched, 100-item sequential replay, learned variants failed required STORE, UPDATE and DUPLICATE actions. A later selective research model chose not to store on all 100 replay cases.

Evidence & limits. A maintainer’s negative engineering report, not a peer-reviewed or independently reproduced result. The replay was fictional and inspected during development; labels were machine-generated, with one model family and seed. The rules baseline also performed poorly and over-stored. The report is dated October 17; the checkpoints are older and the publication timezone was unavailable.

DeepSeek updated its Flash model while keeping the deepseek-flash API alias and pricing unchanged. It reports stronger agentic coding results and recommends DeepSeek Harness version 0.2.1-rc.1 or later.

The release includes MIT-licensed weights. DeepSeek’s reported TerminalBench 4.0 and DeepSWE improvements use different harnesses, so they should be read as separate vendor evaluations rather than a single comparable score series.

Evidence & limits. Vendor-reported results, not independently reproduced here. The changelog is dated October 17; an exact Eastern publication time was not established. There is no Jinja chat template: the reference encoder and deepseek-recipe cover protocol conversion and output parsing, not inference or a tool-execution runtime.

DeepSeek now exposes standalone Search and Fetch endpoints using an existing API key. Search can return page content as well as results, giving builders a direct retrieval layer for their own harnesses.

The documentation lists each standalone service at US$0.75 per 1,000 successful requests. Empty search results, failures and timeouts are not charged. Separately, existing server-side Search and Fetch use through DeepSeek Harness, Claude Code and Codex is currently free; the launch announcement sets the start of per-request billing at October 31, 2026, 16:00 UTC (noon EDT).

Evidence & limits. These are documented prices and announced terms, not measured total workload costs. Vendor claims about search quality and spam filtering have not been independently evaluated here. The standalone API price should not be confused with the later billing start for existing server-side tools.

protoAgent’s maintainer reports that a screenshot-heavy turn left roughly 490 screenshots duplicated across checkpoint states, producing about 87 GB of history. Pruning previously ran only every six hours.

Version 0.208.0 prunes on every write, stores each image once per history database, and adds size alerts. These changes target checkpoint growth rather than merely reducing the model’s context or token budget.

Evidence & limits. A maintainer-reported incident and fix, not independently reproduced here. Existing history may need checkpoint.keep_per_thread adjustment. Stock LangGraph readers need the provided serializer, and downgrading can expose image-reference strings to models.