openclaw - 💡(How to fix) Fix Heartbeat sessions never compact — memory system incompatible with short-session agent patterns [1 comments, 1 participants]

Problem

OpenClaw's memory system relies on session compaction as the trigger for writing summaries to memory. Compaction only fires when a session approaches the model's context limit.

With modern context windows (200K+ tokens), short sessions — particularly heartbeat-driven sessions — may never approach the context limit. A typical heartbeat session might be 500 tokens. With a 200K context window, it would need to run ~400 times before triggering compaction.

This creates a critical failure mode for Paperclip-style agents:

Paperclip heartbeat fires → new short session (500 tokens) → agent does work → session ends
→ repeat every 15 minutes → never hits context limit → never compacts → never writes to memory

Result: Thousands of dollars of agent work, weeks of heartbeat cycles, and zero memory persistence. All that pattern data evaporates when sessions close.

Specific failure modes

1. Heartbeat sessions — designed to be short and frequent, below compaction threshold
2. Paperclip-driven sessions — each wake = fresh session = isolated context = no accumulation
3. High-context agents — with large context windows, even hours of work may not trigger compaction

Current architecture

Memory flow:

1. Session runs → grows toward context limit
2. Near limit → compaction fires → agent reminded to save notes to memory
3. Memory flush → QMD indexes the markdown
4. Short sessions never reach step 1

Proposed solutions (pick one)

Option A: Configurable compaction interval Force compaction after N sessions or N hours, regardless of token count. E.g., compaction: { forceAfterSessions: 50 }

Option B: Heartbeat memory flush A heartbeat-specific memory write that fires on session end without waiting for compaction. Small, consistent, guaranteed.

Option C: Session summary on close Every session writes a mandatory one-line summary to memory on close, even if never compacted.

Why this matters

This affects every Paperclip user running heartbeat-driven agents. The most valuable agent patterns (daily standups, recurring checks, monitoring) are systematically lost to memory. The agent becomes a commodity consumer of tokens with zero institutional memory.

Environment

OpenClaw 2026.4.2, MiniMax M2.7 (200K+ context), Paperclip heartbeat cadence