Safeguarded Compaction

Compaction should keep full history on disk, never split tool-call and result pairs, audit summary quality before replacing history, and flush notes to memory first, with the compaction prompt tuned for recall before precision.

Evidence: strong

Compaction replaces part of a conversation with a summary of it. Done carelessly, it is irreversible, it can corrupt the structure the model relies on, and a poor summary can quietly delete the one detail that mattered. Since most harnesses compact automatically, these risks run on every long session.

Vendor documentation describes a consistent set of safeguards:

  • Keep full history on disk. OpenClaw's auto-compaction summarises older turns as the session nears its limit, or compacts and retries on a context-overflow error, while keeping the full history on disk.
  • Never split a tool call from its result, so the compacted transcript stays structurally valid.
  • Audit the summary before accepting it. OpenClaw's default "safeguard" mode audits summary quality, and a failed audit keeps the original history.
  • Flush notes to memory first. OpenClaw reminds the agent to flush important notes before compacting; Hermes runs a preflight compression check above 50% of context and flushes memory before context is lost.
  • Make lossiness explicit. Hermes routes compression through a pluggable context engine whose default is a lossy summariser, while lossless engines must be selected explicitly.
  • Compact early and in the background. Pi Durable starts background summarisation at a token threshold, makes a request wait for the summary only when close to the window limit, and compacts once and retries if the provider rejects a request as too long.
  • Protect the edges. Meta's multi-turn recipe replaces the middle of the session with a summary carrying the goal, files and plan while keeping the system prompt and recent turns intact, which also respects lost in the middle.
  • Tune the compaction prompt for recall first. Anthropic advises maximising recall on real traces, then improving precision.
  • Log every compaction, so its start, completion and audit outcome are visible afterwards.

These safeguards target a known failure: wholesale rewrites drift towards generic brevity and can collapse, as described in context collapse. They harden the running summary of memory summarization for agents that compact automatically. Keeping full history on disk is the same principle as a session log outside the context window. The broader choice of whether to compact at all is in compaction vs context reset, and where different harnesses sit on memory design is in the harness memory strategy spectrum.

When an agent appears to have forgotten something, the compaction log lines and the audit outcome show whether a compaction happened and whether it was accepted, and the selected context engine tells you whether compression was lossy. An agent that investigates incidents should compact only its working context, while raw logs and traces stay untouched outside the window for later analysis.

Sources

  • OpenClaw — https://docs.openclaw.ai/concepts/compaction
  • Anthropic Engineering, 29 Sep 2025 — https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
  • Nous Research — https://hermes-agent.nousresearch.com/docs/developer-guide/agent-loop
  • Nous Research — https://hermes-agent.nousresearch.com/docs/developer-guide/context-compression-and-caching
  • Earendil Engineering, 1 Oct 2026, background thresholds and compact-once-and-retry — https://earendil.com/posts/pi-durable/
  • Meta Model API cookbook, Agent Patterns, undated (retrieved 3 Oct 2026) — https://dev.meta.ai/docs/cookbook/multi-turn-context-management