feat: evaluate Working Memory retirement if auto-memory proves reliable

[michael-wojcik]

Context

Deferred from PACT v3 Phase D (D1 decision). Tracked for v3.1+.

Description

In v2.8.0, the decision was made to coexist with platform auto-memory rather than retire Working Memory (the CLAUDE.md section synced from pact-memory SQLite). MAX_WORKING_MEMORIES was reduced from 5 to 3 to limit token overlap.

If auto-memory proves reliable over v3.0 usage, consider retiring Working Memory entirely:

• Remove sync_to_claude_md() from working_memory.py
• Remove the Working Memory section from CLAUDE.md template
• Rely on auto-memory for session-to-session context and pact-memory search for structured retrieval

Evaluation Criteria

Before retiring, validate:

1. Auto-memory consistently captures useful session context without manual intervention
2. Auto-memory content is accessible in subsequent sessions (first 200 lines auto-loaded)
3. Removing Working Memory doesn't create a gap in structured context availability
4. pact-memory search remains accessible for deep retrieval (auto-memory can't replace this)

Current State

• Three-layer architecture: auto-memory (broad), Working Memory (structured rolling window), pact-memory search (deep)
• Working Memory is a "view" of the 3 most recent pact-memory entries
• Auto-memory is still relatively new — insufficient usage data to evaluate reliability

Priority

Low — evaluation depends on accumulated v3.0 usage experience.

References

• PR feat: Phase D revised — memory rationalization, agent persistent memory, expanded verification (v2.8.0) #156 (Phase D revised, v2.8.0) — D1 coexistence decision
• pact-plugin/skills/pact-memory/scripts/working_memory.py
• pact-plugin/CLAUDE.md (Three-Layer Memory Architecture section)