[contexto]: Scope mindmaps per OpenClaw agent instead of per instance

[sm86]

Summary

@ekai/contexto should isolate mindmaps per OpenClaw agent by default.

Right now, the plugin flow appears to be instance-scoped: ingestion and retrieval do not seem to carry an explicit agent identifier, so multiple agents on the same OpenClaw instance can end up sharing one Contexto mindmap.

Problem

In multi-agent OpenClaw setups, agents have different roles and histories. If they share the same Contexto mindmap, retrieval can pull in context from the wrong agent.

This causes:

• cross-agent context leakage
• lower retrieval precision
• surprising behavior in multi-agent deployments
• harder debugging when recalled context belongs to another agent

Expected Behavior

Each OpenClaw agent should read and write to its own Contexto mindmap namespace by default.

A query from Agent A should only search context ingested by Agent A. Shared context, if supported, should be explicit and opt-in.

Proposed Direction

• derive a stable agentId or equivalent scope from the OpenClaw runtime
• include that scope in Contexto ingestion payloads
• include the same scope in mindmap search requests
• preserve that scope across buffering, compaction, ingestion, and retrieval
• keep single-agent setups working without extra configuration

Acceptance Criteria

• two agents on the same OpenClaw instance do not contaminate each other’s mindmaps
• searches from Agent A do not return items ingested by Agent B by default
• buffered episodes and compaction preserve the correct agent scope
• existing single-agent installs continue to work unchanged
• docs clarify the default isolation behavior for multi-agent setups

Notes

Relevant code paths:

• packages/contexto/src/engine/utils.ts
• packages/contexto/src/engine/base.ts