Codex Memories enabled and generated locally, but never injected into new threads

Environment

• macOS
• Codex Desktop bundled CLI: 0.140.0-alpha.19
• Standalone Codex installed via official installer: 0.140.0
• Plan: ChatGPT Pro
• Region: Poland / EEA
• Project cwd: /Users/karol/dev/wbijaj

Config

In ~/.codex/config.toml:

[features]
chronicle = true
memories = true

[memories]
disable_on_external_context = false
generate_memories = true
use_memories = true

Observed Behavior

New Codex Desktop threads do not receive any ## Memories section.

A diagnostic prompt in a new thread returns that no Memories were injected.

Expected Behavior

Existing Codex Memories should be injected into relevant new sessions when features.memories = true and memories.use_memories = true, even if Chronicle screen recording is not being used.

Evidence

codex features list shows:

chronicle                            under development  true
memories                             experimental       true

codex app-server daemon version shows daemon running:

{
  "status": "running",
  "managedCodexVersion": "0.140.0",
  "appServerVersion": "0.140.0",
  "cliVersion": "0.140.0-alpha.19"
}

~/.codex/memories/ contains generated memory artifacts:

• raw_memories.md
• phase2_workspace_diff.md
• many files under rollout_summaries/
• Chronicle memory resources under extensions/chronicle/resources/

SQLite state:

• all checked threads have memory_mode = enabled
• stage1_outputs has 34 outputs
• memory_stage1 has many done jobs

However:

• codex debug prompt-input 'Wbijaj create save idempotency PR206 recurrence guardrails: use any relevant Codex Memories if available.' contains no ## Memories
• codex debug app-server send-message-v2 'Diagnostic only: check whether Memories are injected...' produces an assistant reply: No, I did not receive a ## Memories section.
• verbose logging with RUST_LOG=codex_memories_write=trace,codex_core=info during app-server diagnostic did not show any codex_memories_write / Phase 2 events at session start

Local Repair Attempts Already Tried

1. Installed standalone Codex via:

curl -fsSL https://chatgpt.com/codex/install.sh | sh

1. Started daemon:

codex app-server daemon start

1. Confirmed daemon is running.
2. Found stale memory_consolidate_global job and reset it.
3. Disabled retry for old Phase 1 jobs that failed with:

Codex ran out of room in the model's context window

1. Reset memory_consolidate_global.input_watermark to 0.
2. Restarted daemon.
3. Retested prompt injection: still no ## Memories.

Current Relevant SQLite State

• memory_consolidate_global remains in error with retry available.
• input_watermark = 0
• last_success_watermark = 0
• retry_remaining = 3
• stage1_outputs: 34 rows, 0 selected for phase2.

Question

Why is the Memories pipeline not running Phase 2 / not injecting Memories into new Codex Desktop sessions despite enabled config, generated local memory artifacts, enabled per-thread memory mode, and a running app-server daemon?