Environment

• openclaw 2026.4.26
• Plugins enabled: active-memory, memory-core, hindsight-openclaw
• Provider tested: deepseek/deepseek-chat (provider-independent — fails before reaching the model)

Repro

1. Enable both active-memory and memory-core in plugins.entries. No further config beyond enabled: true; defaults for both.
2. Restart gateway. Confirm memory-core loads (openclaw memory --help works; the plugin registers memory_search and memory_get per extensions/memory-core/index.js:789-805).
3. Send any user turn that triggers active-memory's pre-summarizer.

Expected

Active-memory's embedded sub-agent runs against the configured provider with tools memory_search + memory_get available.

Actual

Sub-agent fails with:

[agent/embedded] [tools] No callable tools remain after resolving explicit tool allowlist (runtime toolsAllow: memory_search, memory_get); no registered tools matched. Fix the allowlist or enable the plugin that registers the requested tool.

Followed by:

[agent/embedded] embedded run failover decision: ... stage=prompt decision=surface_error reason=none ... rawError=No callable tools remain after resolving explicit tool allowlist (runtime toolsAllow: memory_search, memory_get); no registered tools matched.

The user-visible chat then shows [assistant turn failed before producing content] (from STREAM_ERROR_FALLBACK_TEXT in compaction-successor-transcript) as a placeholder turn, polluting the saved transcript permanently. Hindsight's auto-recall later quotes that placeholder back as "Prior context" on every subsequent turn.

Apparent root cause

extensions/active-memory/index.js:933 configures the embedded sub-agent with both:

toolsAllow: ["memory_search", "memory_get"],
bootstrapContextMode: "lightweight",

bootstrapContextMode: "lightweight" appears to exclude tools registered by sibling plugins (memory-core in this case), so the runtime tool allowlist filters from an empty set → the error above. Both plugins ship in the bundled extensions directory and are documented to compose, but they don't out of the box.

Suggested fixes (any one would resolve)

• active-memory uses a bootstrap mode that includes memory-kind plugin tools when its allowlist references them.
• memory-core opts into lightweight bootstrap visibility.
• active-memory pre-checks tool availability and skips/logs-warning instead of emitting a placeholder turn that lands in the persisted transcript.

Workaround

Disabling active-memory (plugins.entries.active-memory.enabled: false) stops the placeholder. hindsight-openclaw's autoRecall continues to inject context, so the loss is the bounded summarizer pre-step only. Not a fix — just a way to stop the visible noise.