What happens when sessions build on each other vs start blind. The structural difference between repeating six hours of failure and resolving in five minutes.
Robert says "fix the contact form." Watch what happens in each version of reality.
PUT on snippet → returns 200 but doesn't persist.PATCH → same fake-success behavior.POST → same. Code Snippets API is read-only and you didn't know.MEMORY.md index first.project-master-timeline-mar9-apr13.md. Knows current state of every system.project-cro-form-recurring.md. Three previous sessions documented the same issue..wpforms-form with form#wpforms-form-1234. Two minutes."Trust and productivity diverge sharply over time. Without memory, every session erodes credibility. With memory, every session builds on the last.
Eight categories that show up almost every session. With memory loaded, each becomes a non-event.
| Category | Without Memory | With Memory |
|---|---|---|
| Credentials | Asks Robert each time, breaks flow | Already in reference-api-keys.md |
| Snippet Edits | Tries broken WP API for hours | Knows API is read-only, gives Robert exact line + fix |
| Site Changes | Edits wrong site, hits production | Specifies exact site URL + DB name first |
| Russ Campaigns | Touches them, breaks attribution history | Filters to BSP | ... | Mar 2026 only |
| Daniel ROI | Claims 334x like the dashboard says | Knows it's $0 attributed, says "unverified" |
| Document Overwrites | 4th time blowing away playbook design | Reads existing first, preserves structure, additive only |
| Meeting Context | Asks Robert for recap of yesterday | Pulls Gemini notes from Drive automatically |
| Smart Bidding State | Trusts dashboard data at face value | Knows clean window is 5-7 of last 35 days, flags it |
Six layers, loaded in priority order. Each one feeds the next. Lower layers are deep reference; upper layers are entry points.
Hard rules, mandatory triggers, "protocols" word, intelligence stack endpoints. The constitution.
Master index of every project file. Categorized, linked. The map to the territory.
Current state of every active workstream. CRO form, sewer pages, Daniel AI, ST contract, GSC crisis.
Deep reference: data weapons plan, evolution proof, scientific method engine, gap analysis. SSoT visuals.
pgvector + Claude. Searchable institutional knowledge. Gap: stale on incidents from last 48-72 hrs.
Per-meeting transcripts and action items. Most recent ground truth from Stephanie / Kalen / Audrey calls.
Building the memory was step one. Using it consistently is step two — and the harder one.
Even WITH 31 new memory files, the structural flaw persists if I don't actually use them. Context decays over long sessions. I forget to re-read. I default to "I'll just try the API" instead of "let me check the playbook." The files are useless if they sit unread.
The same six-hour failure pattern returns the moment I skip the index.
MEMORY.md at session start AND every 5–7 tasks
No exceptions. No shortcuts. Read first. Then act.