Prompt Duplication Review

Date: 2026-05-19

Context: Persistent-chat runs with auto-compaction OFF, long sessions (hours/days), and external file-based memory (turn packets, session packets, curated memory). Every turn injects ~1850 tokens of memory context.

Current Per-Turn Prompt Structure

Session-level (set once at ACPX session creation, ~5k cached)

1. Character card (JSON: name, pronouns, personality, scenario, style, goals)
2. Bootstrap prompt (/home/matrix-lite/app/prompts/persistent-chat/bootstrap.md)
3. Agent build prompt from OpenCode config (agent.build.prompt)
4. Agent config (model, temperature, steps, permissions)

Per-turn (injected fresh on every user message)

1. Intro line - "You are Persistent Assistant..."
2. Conversation metadata - key, session_label, room_id, event_ids
3. Memory context - turn-packet.md (lean, ~5 lines since tail removal)
4. Runtime capability context - same hints every turn
5. Recent room-local conversation - full conversation history tail
6. Incomplete/failed turns - prior session failures
7. Attachment metadata - per-message
8. Replied-to context - per-message
9. Batched messages - per-turn
10. Per-turn volatile reminders - 6 identical bullet points every turn
11. User message

Duplication Analysis

1. Recent room-local conversation (ITEM 5)

Duplicates with: ACPX session's retained conversation history (since no compaction)

Why it's there: Designed for sessions that may compact or reload; ensures continuity

Reality: With auto-compaction OFF, the ACPX session never forgets prior turns

Recommendation: Skip injection when session already has history. Only inject on first turn after session creation or compaction.

Savings: Variable (grows with conversation length); at 20+ turns this is substantial

2. Per-turn volatile reminders (ITEM 10)

Duplicates with: Bootstrap prompt rules and character card instructions

Content: 6 bullet points about <matrix_lite_final> tags, tool approvals, etc.

Reality: Identical every turn; already covered in session-level system prompt

Recommendation: Move to session-level bootstrap prompt or OpenCode agent.build.prompt

Savings: ~400 chars / ~100 tokens per turn → ~30k tokens over 300 turns

3. Intro line (ITEM 1)

Duplicates with: Character card + bootstrap prompt (session-level)

Content: "You are Persistent Assistant running through Matrix-lite..."

Reality: This is already the system identity set at session creation

Recommendation: Remove from per-turn prompt; keep in session-level only

Savings: ~200 chars / ~50 tokens per turn

4. Runtime capability context (ITEM 4)

Duplicates with: Itself (same content every turn)

Reality: Static hints to use MCP tools for dynamic runtime info

Recommendation: Move to session-level bootstrap prompt. The content rarely changes

Savings: ~300 chars / ~75 tokens per turn

5. Turn packet "Current thread" section

Duplicates with: Recent room-local conversation (ITEM 5)

Reality: Both summarize the same conversation from slightly different angles

Recommendation: Already trimmed keyword triggers (removed 'test', 'reload'). Consider removing the "Current thread" section from the turn packet entirely since the room-local conversation handles continuity

Savings: ~200-500 chars per turn depending on action items

6. Conversation metadata (ITEM 2) - partially

Duplicates with: Session-level identifiers

Reality: Static fields (conversation_key, session_label) don't change per turn

Recommendation: Move static metadata to session-level; keep only dynamic event IDs per-turn

Savings: ~100 chars per turn

Recommended Changes (Priority Order)

HIGH - Immediate savings, no behavioral impact

1. Remove per-turn volatile reminders → merge into bootstrap prompt (+100 tok/turn)
2. Remove intro line → already in session-level (+50 tok/turn)
3. Move runtime capability context → to bootstrap prompt (+75 tok/turn)

MEDIUM - Architectural, needs testing

1. Skip conversation history injection when session has existing context (variable, biggest saving)
2. Move static metadata to session-level (+25 tok/turn)

LOW - Polish

1. Further trim turn packet "Current thread" or remove entirely
2. Multi-tier decision state (active decisions + archive)

Total Estimated Savings

| Change | Per turn | Over 300 turns |

|--------|----------|----------------|

| Volatile reminders → session | ~100 tok | ~30k tok |

| Intro line → session | ~50 tok | ~15k tok |

| Capability hints → session | ~75 tok | ~22.5k tok |

| Skip conversation tail (non-fresh turns) | ~200-500 tok | ~60k-150k tok |

| Total | ~425-725 tok | ~127k-217k tok |

The conversation tail is by far the biggest saving, but even the smaller items compound to significant numbers over long sessions.