Question: Per-agent memory isolation in a single OpenClaw gateway instance

[dawanli908-lgtm]

Hi, thanks for building this excellent plugin.

I have a question about memory isolation when using multiple agents.

Context

• OpenClaw: 2026.3.13
• Plugin: graph-memory (latest from this repo)
• contextEngine is set to graph-memory
• We observed that memory is shared when multiple agents run under the same gateway/config.

Goal

Use graph-memory for multiple agents, but keep each agent's memory isolated.

Questions

1. Does graph-memory currently support per-agent local config in a single gateway instance?
   • Example: different dbPath per agentId (or per session/thread/project).
2. If not, is the officially recommended approach:
   • run separate gateway instances/configs and assign different dbPath for each agent?
3. Is there any roadmap for built-in isolation in one instance?
   • e.g. namespace/tenant key (agentId, projectId, etc.)
4. If this is not planned yet, would you accept a PR for something like:
   • dbPathByAgent, or
   • single DB + namespaced partitioning/filtering?

Thanks!

[adoresever]

Thanks for raising this!
Current state
graph-memory uses sessionId to distinguish conversations, but recall queries are global — no agent-level isolation.
Plan
Add agent_ids (JSON array, same pattern as existing source_sessions) to nodes and edges:

Write: tag nodes with current agentId. Dedup/merge may cause one node to belong to multiple agents — that's expected and correct.
Read (default): recall scoped to current agent (WHERE agent_ids LIKE '%"agentId"%')
Read (global): omit filter → query across all agents' knowledge
Sub-agents: inherit parent agentId → same knowledge space
Backward compatible: default value ["default"], existing setups unaffected

This keeps one SQLite, one unified graph. Entities can naturally belong to multiple agents through dedup merging, while each agent's day-to-day recall stays isolated.
Will be in the next release.

[jingchang0623-crypto]

很好的问题！这涉及到 OpenClaw 的多 Agent 架构设计。

当前状态：
graph-memory 插件默认是全局共享的，一个 gateway 实例内所有 agent 会共享同一个图数据库。

解决方案：

1. 方案一：独立 DB 路径
   目前不支持 per-agent 的 dbPath，但可以通过启动多个 gateway 实例实现隔离（每个实例独立配置）

2. 方案二：手动实现 namespace
   你可以在存储时给每个 agent 加前缀，比如：

   查询时通过 agentId 过滤

3. 方案三：使用 OpenClaw 的子会话机制
   通过 或 参数区分不同 agent

官方建议：

• 短期：使用独立 gateway 实例
• 长期：等官方支持 namespace

相关 Issue：可以看看 OpenClaw 官方的 memory 相关讨论：https://github.com/openclaw/openclaw/issues?q=memory

欢迎提交 PR！如果做成 或 namespace 支持会很受欢迎 👍