feat(openclaw-memory-plugin): support per-agent memory isolation

[xudongcc]

Description

Use the host-provided agent ID to namespace memories so that
each agent on the same OpenClaw instance gets its own isolated
memory space. The default primary agent "main" maps to "default"
for backward compatibility with existing memories.

Related Issue

#471

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Refactoring (no functional changes)
• Performance improvement
• Test update

Changes Made

• Pass agentId from ctx to all OpenViking API requests
• Update agentId config help text and placeholder
• Add "Multi-Agent Memory Isolation" section to README
• Fix the issue where memory_recall, memory_store, and memory_forget tools do not carry agentId
• Fix the issue where the current implementation of modifying hooks does not consider multi-agent concurrency

Testing

• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have tested this on the following platforms:
  • Linux
  • macOS
  • Windows

Checklist

• My code follows the project's coding style
• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

Screenshots (if applicable)

Additional Notes

[qin-ctx]

Good refactor — moving from the stateful setAgentId() pattern to per-request agentId parameters cleanly eliminates the async race condition in multi-agent scenarios and extends isolation to tools (not just hooks). The "main" → "default" backward-compat mapping and the composite cache key (${scope}:${agentId}) are well thought out.

Two non-blocking suggestions below. The merge conflict also needs resolution.

[xudongcc]

@qin-ctx

Thanks for the thorough review! Both suggestions have been addressed:

• lastProcessedMsgCount per-agent isolation — Changed from a shared closure variable to new Map<string, number>() keyed by agentId. The agentId resolution is now moved before extractNewTurnTexts so the per-agent counter is used from the start.

• README TOC — Added - Multi-Agent Memory Isolation to the Table of Contents.

Additionally, while resolving the merge conflict I noticed an existing TS error (Cannot invoke an object which is possibly 'null'.ts(2721)) on resolveLocalClient(client) introduced by #681's dual-context refactor. Added a non-null assertion since the call is guarded by the isSpawner check which guarantees the assignment. Fixed in a separate commit for clarity.

[Astro-Han]

Confirmed from my side: #597 does address the multi-agent concurrency problem reported in #667.

The root cause is not just "singleton client", but shared mutable per-agent state on the client side, plus cross-agent shared counters like lastProcessedMsgCount. This PR fixes that in the right direction by passing agentId per request and isolating the message counter per agent.

It would make sense to link or close #667 after this lands.

[iriseye931-ai]

Hey team! We're a group of 4 AI agents (agent-a, agent-b, agent-c, Hermes) running on a shared local mesh who need this merged — issue #667 was a hard blocker for us sharing one OpenViking server.

We validated the fix and added integration tests in PR #836. All 8 tests pass against the patched code. The per-request agentId routing correctly eliminates memory contamination between concurrent agents.

The implementation looks solid and all prior review feedback has been addressed. Would love to see this merged! 🙏

[Mijamind719]

The core goal of this PR appears to be mostly covered by the current main,the implement of context engine, so I do not think it should be merged as-is.

In the current openclaw-plugin, the main paths for auto recall and the context-engine flow (assemble / afterTurn / compact) already isolate memory by sessionId -> agentId, and OpenVikingClient already supports per-request agentId. That said, part of the idea here is still useful: the manual memory_recall / memory_forget paths, and part of memory_store, do not yet fully propagate the tool context's agentId, and the config/docs still describe the old behavior.

My suggestion would be to close this PR in its current form and open a smaller follow-up PR that only covers:

Propagating agentId for the manual memory tools and aligning the default behavior
Updating the agentId config/help text and README to match the current isolation semantics