Silent turns and the memory flush

Silent housekeeping (NO_REPLY)

OpenClaw supports "silent" turns for background tasks where the user should not see intermediate output.

• The assistant returns an entire final that trims to uppercase NO_REPLY to request no user-visible text. Lowercase, punctuation, quotes, JSON, and trailing tokens remain visible.
• The terminal producer records the decision and prepares output once; delivery and clients consume that fact instead of parsing text. One literal leading NO_REPLY\n removes only that control line and preserves the remainder as visible text.
• Streaming control buffering belongs to producer preparation, so a possible control line does not leak into drafts or typing activity.
• Use this for intentional silence, not missing output. External turns honor it even when a reply was required; internal result-required turns still need an answer.

Pre-compaction memory flush

Before auto-compaction happens, OpenClaw can run a silent agentic turn that persists durable state through the selected memory provider so compaction cannot erase critical context. Memory Core writes memory/YYYY-MM-DD.md in the agent workspace; other providers can expose their own persistence tools. OpenClaw monitors session context usage, and once it crosses a soft threshold below the compaction threshold, it sends a silent "write memory now" directive using the exact silent token NO_REPLY so the user sees nothing.

Memory flushing runs against a private, detached view of the conversation. Its internal prompts and replies never enter the original transcript, including when a new user message interrupts it. Memory-file writes remain durable. Required preflight excludes the already-admitted waiting user; post-reply flushing includes the completed turn. Any compaction inside the flush affects only its private view; the original conversation has a separate compaction step.

Config (agents.defaults.compaction.memoryFlush), full reference at /gateway/config-agents:

The host resolves the reserve, soft margin, and transcript-byte guard from this config and the active context window before it asks the selected memory provider for flush content and persistence. Providers inherit those values unless they deliberately override an optional timing field in their plan. When enabled is false, the host skips the provider resolver. Without a flush plan, maintenance still uses the effective compaction reserve.

The default 20,000-token reserve is capped at one quarter of the context window. The default 4,000-token soft margin is capped at half the window left after that reserve. For a 32,768-token window, the host therefore uses an 8,192-token reserve and a 4,000-token soft margin. Early flushing starts at 20,576 projected tokens. Blocking token compaction starts at 24,576, or later if an applicable server threshold is higher. Between those thresholds, memory flushing can run without requiring compaction. The transcript-byte guard defaults to 2 MiB and remains independent. Nonpositive thresholds suppress their trigger.

When memory flush refreshes stale usage, it includes projected messages appended after the latest valid provider usage report before saving the total as fresh. The following compaction check therefore accounts for that later transcript growth.

Notes:

• The built-in prompt and system prompt include a NO_REPLY hint to suppress delivery.
• The flush uses the active session model and its normal fallback chain. Use a local session model and local fallbacks to keep memory extraction local.
• The flush runs once per compaction cycle (tracked in the session row).
• The flush runs only for embedded OpenClaw sessions; CLI backends and heartbeat turns skip it.
• Memory Core's file flush is skipped when the session workspace is read-only (workspaceAccess: "ro" or "none"). A tools-arm provider does not require writable workspace access.
• See Memory for the workspace file layout and write patterns.

OpenClaw exposes a session_before_compact hook in the extension API, but the flush logic above lives on the Gateway side (src/auto-reply/reply/memory-flush-plan.ts, src/auto-reply/reply/memory-flush.ts, and src/auto-reply/reply/agent-runner-memory.ts), not on that hook.