[Feature]: (memory): optional PowerMem retrieval for system prompt context

[Teingi]

What workflow is missing or painful today?

Motivation

OpenHarness memory today is Markdown-first: MEMORY.md plus topic files under the project memory directory, with keyword/heuristic relevance via find_relevant_memories. Teams using PowerMem (vector + full-text + graph, optional intelligent memory features) have no first-class way to inject retrieved memories into the harness system prompt.

Proposed change

Add an optional integration layer that:

• Keeps default behavior unchanged (memory.backend: local).
• Supports HTTP against a running PowerMem API server (POST /api/v1/memories/search).
• Optionally supports in-process SDK via optional extra openharness-ai[powermem] (requires Python compatible with the powermem package, typically 3.11+).
• Supports hybrid: existing local Markdown relevance plus PowerMem retrieval.

Alternatives considered

No response

Additional context

No response

[Teingi]

powermem: https://github.com/oceanbase/powermem

[siaochuan]

Great timing — we've been working on exactly this problem from a different angle.

Our approach: Dongtian

Dongtian is a lightweight MCP-based memory server (SQLite + FTS5 + optional embeddings) that already integrates with OH. It ingests sessions from Claude Code, Codex, OpenCode/DeepSeek, and ChatGPT into a structured palace (Wing → Room → Drawer), with hybrid search and knowledge graph extraction.

We've tested it end-to-end with OH + MiniMax — the agent can search across 16k+ drawers from multiple AI tools and remote machines via SSH sync.

Suggestion: define a memory provider interface

Rather than coupling OH to one specific backend, I'd propose defining a MemoryProvider protocol that both PowerMem and Dongtian (and future backends) can implement:

class MemoryProvider(Protocol):
    async def search(self, query: str, limit: int = 10) -> list[MemoryResult]: ...
    async def store(self, content: str, metadata: dict) -> str: ...
    async def close(self) -> None: ...

Configuration in settings.json:

{
  "memory": {
    "backend": "local",           // default: unchanged
    "providers": {
      "powermem": {"url": "http://localhost:8080"},
      "dongtian": {"command": "python -m dongtian"}
    }
  }
}

This way:

• Default behavior unchanged (backend: local)
• PowerMem: HTTP provider via /api/v1/memories/search
• Dongtian: MCP provider (already works, see PR #65)
• Hybrid: merge results from multiple providers with weighted scoring

Related PRs

• #65 — auto-extract local environment rules from sessions (personalization layer)
• #63 — fix MCP tool type mapping (needed for MCP memory servers to work with non-Anthropic models)

Happy to collaborate on the provider interface design. We can contribute the MCP/Dongtian adapter and tests.