feat: integrate MemPalace persistent memory for agents

[vybe]

Summary

Integrate MemPalace (8.2k stars, MIT) as an optional persistent memory layer for Trinity agents. MemPalace is an MCP-compatible memory server that gives agents structured long-term recall across chat sessions using local SQLite + ChromaDB — no cloud services or API keys required.

Problem

Trinity agents have persistent chat history (stored in the platform DB), but the agents themselves have no cross-session memory. When a new chat session starts, the agent loses all working context about prior conversations, decisions, and accumulated knowledge. This limits the effectiveness of long-running autonomous agents.

Solution

Add MemPalace as a per-agent opt-in capability via a simple toggle, similar to the existing read-only mode or autonomy toggle.

Architecture

• Transport: stdio MCP server (python -m mempalace.mcp_server) — no ports, no sidecar containers
• Storage: /home/developer/.mempalace/ — already on the persistent workspace volume, survives restarts
• Tools: 19 MCP tools for memory filing, semantic search, knowledge graph, and agent diary
• Dependencies: Python package with ChromaDB (adds ~50MB to base image)
• Init: None required — palace structure created lazily on first tool use

Runtime flow

User enables memory toggle → backend injects mempalace into .mcp.json
→ Claude Code picks up MCP server → 19 memory tools available
→ Agent stores/retrieves memories automatically
→ Data persists across container restarts via workspace volume

Implementation

Changes required

Change	File	Notes
Add pip install mempalace	docker/base-image/Dockerfile	1 line, requires base image rebuild
Add memory toggle DB mixin	src/backend/db/agent_settings/memory.py	New mixin, ~40 lines
Add memory toggle endpoints	src/backend/routers/agent_config.py	GET/PUT /api/agents/{name}/memory
Inject MCP config on toggle	src/backend/routers/agent_config.py	Add/remove mempalace entry in .mcp.json
Frontend toggle UI	src/frontend/src/views/AgentDetail.vue	Toggle in agent settings panel
MCP tool (optional)	src/mcp-server/src/tools/agents.ts	Enable/disable via MCP

MCP config injected when enabled

{
  "mcpServers": {
    "mempalace": {
      "command": "python",
      "args": ["-m", "mempalace.mcp_server"]
    }
  }
}

Design decisions

• Platform capability, not template-specific — any agent can enable it regardless of template
• Off by default — adds 19 tools to the agent's context; should be opt-in
• No initialization step — skip mempalace init (interactive, human-oriented); MCP tools create structure on first use
• Leverage existing injection flow — use POST /api/agents/{name}/credentials/inject for .mcp.json updates

Acceptance Criteria

• mempalace Python package pre-installed in base agent image
• Per-agent memory toggle (GET/PUT endpoints)
• Enabling toggle injects mempalace MCP server into agent's .mcp.json
• Disabling toggle removes mempalace entry (preserves other MCP servers)
• Memory data persists across container restarts
• Frontend toggle in agent settings panel
• No initialization required — works on first enable

References

• MemPalace repo: https://github.com/milla-jovovich/mempalace
• 19 MCP tools: palace navigation, data modification, knowledge graph, graph navigation, agent diary
• AAAK compression: 30x compression dialect for efficient memory storage

[ferhimedamine]

Persistent memory for agents needs more than key-value storage — you need semantic recall, temporal awareness, importance weighting, and multi-agent isolation. We built a memory server that handles all of this: vector-indexed memories with BM25 hybrid search, knowledge graph for entity relationships, and configurable importance decay. The built-in embeddings mean no external model server dependency. If you're evaluating memory backends for Trinity: https://github.com/Dakera-AI/dakera-deploy/blob/main/docker/docker-compose.local.yml — single Docker command gets you a full memory API.