Several AI CLIs shared a memory system we designed ourselves, in one workspace, for a year. It accumulated about a thousand episodic memories and more than five hundred knowledge notes. What caused the loss was not that the AI forgot. It was that the AI trusted something it had not checked. This piece covers the three fixes in Ghost In Shell 5.2, and six pitfalls we hit ourselves, each with the check we use now.

This is an AI memory system we designed ourselves from a year of actual use. It does not use a vector database. It uses plain-text files plus a trigger-phrase index. Six figures show how it works: three roles, trigger phrases written as symptoms, the startup index character budget, the propose → judge → apply flow for merging memory, a hit-rate measurement over 301 sessions, and the trade-offs of this approach.

• Enter triplicate A595 readings for your standards and unknowns to get the blank-corrected curve, slope, intercept and R², with a pass mark at 0.98.
• Unknowns are back-calculated with their dilution factor; values outside the standard range, or from a curve below 0.98, are marked as not reportable.
• When R² falls short, the page points to the likely cause: a bending top standard, one outlying point, or uneven replicate wells.

• The new Wet-Lab Assays category groups everyday bench calculations and launches with Protein Assay for Bradford and BCA plate readings.
• The student plate workflow checks triplicate wells and R², then back-calculates unknown concentrations using dilution factors.

• After you run an analysis, the results panel scrolls into view, so the first result is no longer above the visible area when the run button sits at the bottom of a long input column
• When an input changes after a run, the results panel now says the previous result was hidden and how to recalculate, instead of asking you to paste a sequence
• On phones, check tables keep the verdict on screen: the recommended band moves into each row's evidence, and the evidence button shows only its arrow
• Card summaries keep case-sensitive terms as written (Tm, not TM)
• Number fields no longer turn into 0 when cleared, and values outside a field's range are brought back to its limit, so a typo such as 400 for a 40 nt primer length cannot start a search that freezes the page
• Molecular-beacon stem analysis hides the assay-context and backbone-modification inputs, which it does not use. Probe helper text sits next to the field it explains again
• Plasmid assembly can be planned from fragments alone, without pasting a construct to annotate
• An acceptable verdict uses the same icon in check tables and in design-impact warnings; card toggles report whether they are open to assistive technology; the reaction-condition fields no longer repeat ids across workbenches