openclaw - 💡(How to fix) Fix Feature: allow per-agent/runtime restriction of memory_search allowed corpora

Summary

Please add an official per-agent/runtime policy to restrict which corpora the memory_search tool exposes and accepts.

Our concrete need is to run operational live-chat agents with memory_search available for corpus=memory, while hiding or rejecting corpus=sessions at the tool-schema/tool-policy level.

This is related to #49524, but slightly different: #49524 is about fail-open liveness when memory_search stalls. This request is about preventing sensitive or high-risk corpora from being exposed to live sessions/subagents in the first place.

Environment

• OpenClaw: 2026.5.27 stable
• Agent: main
• Channel: Telegram DM, used as a live operational cockpit
• Auth/model route: OpenAI Codex/OAuth
• openclaw memory status --deep: Sources: memory, Dirty: no, embeddings/vector/FTS ready
• memorySearch.extraPaths: intentionally limited to workspace docs/workflows
• agents.defaults.memorySearch.sources: not present as an explicit config path in this install
• experimental.sessionMemory: not present as an explicit config path in this install

Observed behavior

Even when the runtime memory status resolves to Sources: memory, the memory_search tool still exposes/accepts corpus=sessions.

In our live Telegram context, a read-only probe with corpus=sessions stalled until an external 90s timeout and caused observable latency/memory pressure. The gateway and memory index were healthy afterward, but the surface remains risky.

We also validated:

openclaw memory status --deep
# Sources: memory
# Dirty: no
# embeddings/vector/FTS ready

openclaw config get agents.defaults.memorySearch.sources
# path not found in this config

openclaw config get experimental.sessionMemory
# path not found in this config