[Feature] Memory calibration agent — validate thread memory against project config

[davesleal]

Problem

Thread memory persists stale or incorrect context across threads. When the agent hallucinates during one conversation (e.g., calls a web app a "macOS weather app"), that gets saved to thread memory and poisons every future conversation in that channel.

Currently the only fix is manually deleting .shellack/thread-memory/*.md.

Solution

A memory calibration agent (Haiku) that runs on every thread memory load:

1. Reads the thread memory
2. Compares against the project config (name, language, platform, description from projects.yaml)
3. Flags or removes contradictions
4. Returns only validated context

Flow

New thread starts → load thread memory → calibration check:

Haiku:
  "Does this thread memory match the project config?
   Project: Atmos, platform: web, language: typescript
   Memory says: macOS app, Swift
   → INVALID. Discard memory, start fresh."

Calibration rules

• If memory references a different platform than config → discard
• If memory references a different language than config → discard
• If memory contains open questions that STATE.md already answers → resolve them
• If memory is older than 24h → flag for review, prefer STATE.md
• If memory aligns with config → pass through

Cost

One Haiku call per new thread (~$0.0005). Only runs when thread memory exists. Skip if no thread memory file found.

Alternative: TTL-based expiry

Simpler approach — thread memory files expire after a configurable TTL (e.g., 24h). Stale files are deleted on load. No Haiku call needed.

Could combine both: TTL for age-based cleanup + calibration for content validation.

Priority

High — this is actively causing bad UX. Every wrong answer gets persisted and repeated.