Public Evidence Lab / Dossier O-3448

Corroborated▼ FadingOPPORTUNITY DOSSIER · O-3448

AI failures to retain instructions across sessions

Collected sample: 8 independent content items across 2 source families. Known publication dates: 2026-08-20 — 2026-10-22.

First seen 2026-08-20 · Last updated 2026-11-19 · Recalculated daily

27.8
Evidence confidence, not a return forecast
8
Independent source items, deduplicated by post
2
Public source families
1
Payment evidence: current spend or explicit intent
Product hypothesisPersistent Instructions Across AI Sessions (unvalidated; this site does not provide this tool)
ProblemAI assistants often lose process conventions when a new session begins, forcing users to retrain the same behavior. This drift is especially disruptive when users rely on persistent search rules or documented workflows such as AGENTS.md.
People affectedIndividuals and small teams who use AI assistants for structured information retrieval, content creation, or coding and depend on stable workflows. The issue particularly affects operators, analysts, and developers with recurring conventions.
Named alternativesClaude, Cursor
Topicsai-assistantsinstruction-memorysession-continuityworkflow-consistencypersonalization
Collected sample dates · 12 weeks
08-3109-2810-2611-16

These observations describe collected samples, not platform-wide prevalence; they cannot establish whether a problem is widespread or accelerating. Collection scope changed on 09-16 and 09-27; monthly counts must not be read as trends.

Representative evidence

Full evidence chain · 8 items · 8 representative excerpts shown
Product complaint★★★★★Current spend

Post title: Claude losing its mind?

I have been using Claude for several months (paid)…

The public layer keeps only a minimal excerpt. Open the source for full context.

Reddit2026-08-29View source ↗
Pain★★★★☆First-hand pain

Post title: My AI agent remembers everything. Until I start a new session.

I have an AI agent who memorizes facts during one …

The public layer keeps only a minimal excerpt. Open the source for full context.

Reddit2026-10-22View source ↗
Pain★★★★☆First-hand pain

Post title: Keeping an AI coding agent consistent across sessions — what actually works?

Process drift between sessions. A convention gets …

The public layer keeps only a minimal excerpt. Open the source for full context.

Reddit2026-08-20View source ↗
Pain★★★☆☆First-hand pain

Post title: Questions??

I use AI a lot and one thing that keeps pissing me…

The public layer keeps only a minimal excerpt. Open the source for full context.

Reddit2026-10-06View source ↗
Need★★★☆☆Product gap

Post title: Memories in Cursor

When I was using Codex, I used AGENTS.md for the c…

The public layer keeps only a minimal excerpt. Open the source for full context.

Reddit2026-08-31View source ↗
Pain★★★☆☆First-hand pain

Post title: Help with food categorisation project

I constantly have to train the AI to search based …

The public layer keeps only a minimal excerpt. Open the source for full context.

Reddit2026-08-20View source ↗
Product complaint★★★☆☆First-hand pain

Post title: Claude stubbornly refusing to do something it would happly do in a previous session

Then when i switched over to a new session, sudden…

The public layer keeps only a minimal excerpt. Open the source for full context.

Reddit2026-08-20View source ↗
Need★★☆☆☆First-hand pain

Post title: Show HN: Lossless-memory – a personal AI memory that never summarizes

Is the date really useful? Usually I want the AI to remember rules, like "always do X before committing Y" and a timestamp doesn't change anything there.

Hacker News2026-09-21View source ↗

What this evidence does not show yet

To reach Payment-backed we still need:

  • renewed mentions — the recent trend is below the promotion bar

Follow how this opportunity evolves

Every item traceable · free · we email you only when this market actually changes

Scoring summary

Evidence confidence = strength × evidence volume × source diversity × payment × competition × trend × 10. Counts use independent content items; only first-hand pain, current spend, explicit willingness to pay and concrete feature requests affect the score. Repeated posts by one author are discounted. Scoring and status changes follow deterministic rules.

“Payment-backed” means first-hand payment evidence exists in the record. It does not mean the business is worth building. “Fading” is a recency tag shown alongside any evidence level, not a lower level. Read the full methodology.