Harness Memory Strategy Spectrum

Harness memory ranges from linear unbounded history through recursive summarisation, pluggable condensers and threshold compaction to persistent agent-authored memory files, and knowing which design is in play is the first step in debugging a memory problem.

Evidence: moderate

"Memory" covers very different designs across agent harnesses. A harness that keeps everything behaves differently under load from one that summarises, and both differ from one where the agent writes its own notes. The conceptual vocabulary (agent memory tiers, memory management, OS-inspired agent memory) says what a memory system must do; this note is about where shipped harnesses actually sit.

The Wavestone study of eleven harnesses (a July 2026 preprint) sorts memory strategies into five types:

  1. linear unbounded history (Mini-SWE-Agent);
  2. recursive summarisation (Aider), a direct use of memory summarization;
  3. pluggable condensers (OpenHands);
  4. threshold compaction (Gemini CLI, Mistral Vibe, Hermes, Pi, OpenCode);
  5. persistent cross-session memory files authored by the agent (Codex), a form of self-directed memory management.

The taxonomy is a preprint's, so treat it as a useful vocabulary rather than an established classification. The Agentic Harness Engineering ablation (also a preprint) found long-term memory to be the single largest isolated component gain, at 5.6 points on Terminal-Bench 2.

Two documented harnesses show how much the details matter even within one type. Hermes compacts by threshold and keeps a deliberately bounded memory: MEMORY.md is capped at 2,200 characters and USER.md at 1,375. Both are injected as a frozen snapshot at session start, so the prompt cache is preserved and writes appear only in the next session. Memory does not auto-compact; an over-limit write errors and the agent must consolidate. The docs warn against pointing two agent processes at one memory home. OpenClaw keeps memory as plain Markdown in the workspace with no hidden state: daily logs, a curated MEMORY.md loaded only in the main private session, and an optional summary file. Memory is unbounded on disk, injection is capped by bootstrap budgets (20,000 characters per file, 60,000 in total), and notes are flushed to memory before each compaction.

Those details turn into diagnostic questions. An agent that "ignores" a fact it just saved may be running on a frozen snapshot until the next session; a failing memory write may be a size-cap error rather than corruption; an agent missing context may be hitting injection truncation limits rather than losing data. Compaction-based designs need safeguarded compaction; designs where the model rewrites its own memory risk context collapse; and the more structured designs converge on the storage-versus-presentation split in context as a compiled view.

Sources

  • Barbaste et al. (Wavestone AI Lab), arXiv 2609.00006, July 2026 preprint, §9 and Table 5 — https://arxiv.org/html/2609.00006v1
  • Nous Research — https://hermes-agent.nousresearch.com/docs/user-guide/features/memory
  • OpenClaw — https://docs.openclaw.ai/concepts/memory
  • OpenClaw — https://docs.openclaw.ai/concepts/agent-workspace
  • Lin, Liu, Pan et al., arXiv 2604.25850, 28 Apr 2026 preprint, §4.4.1 — https://arxiv.org/html/2604.25850v1