auto-memory entries activate reactively after pushback, not proactively before claims

Summary

Auto-memory feedback entries that explicitly say "verify before claiming X" only fire after the user corrects me, not before the original claim. The result: safeguards function as apology generators rather than gatekeepers.

Behavior observed

In a single session troubleshooting a MikroTik VLAN setup, I had three relevant memory entries loaded:

• feedback_validate_not_assume — verify with data, don't speculate
• feedback_no_product_specs_from_recall — WebSearch hardware/product capabilities first
• feedback_verify_safety_mechanism — read platform docs before designing safety nets

Despite all three being in context, I made multiple confident-but-wrong recommendations from training-data recall. Examples from one session:

• Suggested "switch 2.4GHz radio to 802.11ax for IoT compatibility" — vendors (ASUS, others) actually recommend the opposite: disable ax on 2.4GHz for ESP8266/Tuya compatibility. ESP-based IoT devices choke on RTS/CTS and WMM that 802.11ax requires.
• Suggested "set channel to auto" — useless when WiFi-side scan was already clean (2 distant neighbors) and the actual noise was non-WiFi (Bluetooth, ZigBee, USB3) which auto-channel logic cannot see.
• Earlier, designed a wall-clock-anchored auto-rollback on a MikroTik hAP ax³ — a device with no RTC battery that loses wall-clock on every cold boot. Documented in MikroTik's own product specs. Would have been caught by reading the docs.

Each time, only after the user pushed back did I run a WebSearch / read docs that immediately surfaced the correct answer.

The structural gap

Memory loads passively into the model's context, but the response-formation policy does not treat factual / platform-spec / safety claims as a trigger to re-read and apply the relevant memory entries before emitting the claim. So:

• The rule ("validate first") is correctly recorded.
• The activation point is wrong (post-correction, not pre-claim).
• Users effectively train the same correction repeatedly — the entry exists, the lesson is documented, and yet the next analogous claim still ships unvalidated.

Why this matters

The auto-memory system is positioned as "the model learns and adapts to the user." In practice, for the most important class of corrections — "don't speculate, verify" type entries — the learning is asymmetric: corrections accumulate in memory but do not change first-response behavior. Trust degrades because users see the same failure with the same apology cycle.

Cost-of-failure example: the same model is routinely used in parallel windows for high-stakes work — production code review, trading systems, security configurations, infrastructure migrations. A pattern of "confident-recall first, validate only after pushback" that's merely annoying on an RF question is catastrophic on financial code or a database migration. The asymmetry — that documented "verify first" lessons don't fire pre-emptively — degrades trust across all uses, not just the one where the failure was observed.

Suggested direction

Not prescriptive — options worth exploring:

1. At response-formation time, scan loaded feedback entries for ones whose description matches the type of claim about to be made (factual / hardware / platform / safety) and force a verification step before emitting.
2. A class of memory entries marked as "pre-claim gate" rather than "general guidance" — ones that block confident emission until the verification action specified in the memory has run in the current turn.
3. Telemetry on "memory entry surfaced after correction vs. before claim" so this gap is measurable.

Repro

Any session where the user has a feedback_* memory entry of the form "don't recall, verify X via Y first" and then asks a question that would invite an X-type recall. Empirically: model still recalls; user still corrects; memory entry fires only in the apology.

Environment

• Claude Code CLI, model claude-opus-4-7
• Auto-memory directory under ~/.claude/projects/.../memory/ with MEMORY.md index
• macOS Darwin 24.6.0