openclaw - 💡(How to fix) Fix Active Memory: add single-shot mode (no embedded agent loop) for low-latency preflight injection [1 participants]

Summary

Request a new plugins.entries.active-memory.config.mode: "single-shot" (or similar) that bypasses runEmbeddedPiAgent and produces the memory summary in one LLM round-trip + one qmd query, instead of an N-turn agent loop.

This is the second of the two fixes proposed in #72347. That issue addressed the timeoutMs enforceability bug; this one is the latency-architecture sibling.

Proposal

Roughly:

// pseudocode for AM single-shot path
const queryText = deriveQuery(message, config.queryMode);              // existing
const hits = await qmd.search(queryText, config.qmd);                  // ONE search
const prompt = buildSingleShotPrompt(message, hits, config.promptStyle);
const reply = await llm.complete({                                     // ONE API call
  model: config.model,
  prompt,
  maxTokens: chars2tokens(config.maxSummaryChars),
  abortSignal: budgetSignal,
});
return parseNoneOrSummary(reply);

No tools, no agent loop, no runEmbeddedPiAgent. Wall-clock ≈ 1 API round-trip + 1 qmd search.

Why this is needed in addition to the hard timeout from #72347

A hard timeout caps the worst case but doesn't change the mean. Empirically, the AM tool loop runs 6–12s end-to-end on fast models (gpt-5.4-mini direct API benchmarks at ~1.8s per call; multiply by 3–4 turns). To make AM useful as a preflight (not just a brake), the architecture itself needs to be lighter for promptStyle: "preference-only"-style use cases.

Scope (intentionally narrow)

• In scope: new opt-in mode config value, single API call, no tool use, returns NONE or a ≤maxSummaryChars summary.
• Out of scope: changing the existing multi-turn mode (which makes sense for recall-heavy / contextual styles), or auto-detecting which mode to use.

Repro showing why current architecture is too slow

(Same setup as #72347 — single Telegram DM to the agent, AM enabled. Shipped logs show 12–22s elapsed across 4 models, including local ollama and direct OpenAI API. Direct API benchmark of gpt-5.4-mini: 1.82s. Bottleneck is the loop, not the model. Full table in #72347.)

Note

Filing as a separate ticket per the suggestion in the closing comment of #72347. Single-purpose, hopefully easier to triage and not at risk of being auto-closed by the same bot pattern that mis-closed the parent.