ChatGPT, Anthropic, Devin, Honcho, OpenClaw and Letta now run a second memory writer while you are away. What it reads, what it deletes, what it adds that nobody said, and what none of them have measured. An illustrated page, as of 8 October 2026.
On 600 comparison questions with one passage missing, thinking abstained and direct answering named whichever option the passage described. Which passage was kept moved DeepSeek V4.1 Flash's thinking gain from -79.1 to +10.7 points, and two added sentences removed the deficit.
How I built LAPSE, a synthetic benchmark for LLM memory: 15 hand-written prompts and 66 hand-read probe calls, minimal-pair templates, invented names checked against corpora and the web, dates computed from the run date, and a smoke test read by hand.
I pulled every score ACL Rolling Review has published: 69,781 scored submissions across 35 cycles, 2021 to 2026. No paper ever averaged a 5.0. The scores are per-paper aggregates, not individual reviews, the mode is 2.5, and the meta score swings wider than a paper's averaged reviews.
The workflow I use across concurrent paper projects: each reported value points to a scorer and run, stop conditions precede spending, full runs require five checks, logs are append-only, and detector outputs receive a manual audit.
Across 20 models, a comment typed after a bare paste ends up inside the edited text at rates from 7.7% to 66.7%. A blank line does not help; boundary markers reduce absorption in 19 of 20 models.
Across five models, a nonbinding monitor improved recovery from corrupted tool results when its receipt named callable alternatives. Detection alone did not produce the measured gain.
Memory writers store "I'm staying in Pasadena" as "lives in Pasadena." A one-model pilot, then the LAPSE paper: three writers flattened the progressive in 244 of 381 pairs, never the reverse, across 11 configurations and in mem0, Graphiti, and Letta.
Part one replaced a 3.4 MB video with 40 KB of DOM animation. Part two adds a camera: zoom into the action, pan to follow the cursor, pull back for scene changes. Plus the SEO, accessibility, and agent arguments I skipped the first time.
A public companion to our arXiv paper on LLM memory evaluation: the same ranked list can produce different winners when Raw, Source, and Canonical scoring targets are swapped.
A 15-second hero video was 3.4 MB, letterboxed on phones, and ignored prefers-reduced-motion. I replaced it with a scripted GSAP animation: one timeline, cursor choreography, scene transitions, under 40 KB gzipped.
A protocol proposal for voice-agent interruptions: generated audio races ahead of playback, current stacks expose cancellation and truncation, and agents still need first-class heard state and resumable threads.
On 500 HotpotQA questions, one stored RECOMP output raised mean exact match across 20 readers from 33.0% to 40.4% and shrank the gap between the lowest and highest raw scorers from 31.8 points to 7.8. Companion to the paper Compression Is Not Evaluation-Neutral.
An implementation report for playimposter.xyz covering page-specific metadata, canonical URLs, multilingual hreflang clusters, structured data, crawl directives, sitemaps, and server configuration.
Part 1 of my deepseek-go series: why I built it, the production problems it solved, and the architecture principles that shaped the package.
Part 2 dives into deepseek-go internals: client options, request/response handlers, streaming, FIM, JSON mode, and provider compatibility.
Part 3 analyzes adoption and growth with live GitHub stats, release cadence, star-history graphs, and what those trends changed in my engineering priorities.
Part 4 closes the series with contributor thanks, hard SDK lessons, and my roadmap for deepseek-go's next stage.
Part 5 details my Go testing strategy for deepseek-go: layered tests, stream safety checks, and release confidence practices.
Part 6 explains how I integrated Ollama into deepseek-go with explicit adapters, stream resilience, and practical provider compatibility.
At Prediction3d, edge deployment made data quality, memory use, latency, quantization, pruning, and exported-runtime checks part of the model-selection decision.
Part 3 of the Go web services series: testing pyramid, deployment safety checks, rollout strategy, scaling sequence, and production checklist.
Part 2 of the Go web services series: repository boundaries, middleware ordering, authentication, database handling, and observability.
Part 1 of a detailed Go web services series: setup, server lifecycle, configuration, error handling, and Python-to-Go workflow.
How I designed a paper-aesthetic texture and integrated it into a modern React portfolio using Tailwind CSS.
