Mnemos Memory

Mnemos Memory

Mnemos is a local-first memory layer for coding agents. Use this skill to guide users or OpenClaw agents onto the supported install path, explain the operating loop, and keep compatibility claims accurate.

Default path

• Prefer pip install "mnemos-memory[mcp]" and mnemos ui.
• For OpenClaw / ClawHub, teach the agent to self-install mnemos-memory[mcp], run mnemos ui, then wire mnemos-mcp to the canonical MNEMOS_CONFIG_PATH before relying on memory.
• Recommend SQLite as the starter store unless the user already has a scale reason for Qdrant or Neo4j.
• Recommend a real embedding provider (openclaw, openai, openrouter, or ollama) for production retrieval quality.
• Validate setup with the control-plane smoke check or mnemos-cli doctor.

Claim discipline

• Safe to claim: local-first scoped memory, MCP tools, SQLite starter profile, Claude Code plugin flow, documented Codex flow.
• Be explicit that deterministic auto-memory is shipped for Claude Code via hooks.
• For Codex, Cursor, OpenClaw, and generic MCP hosts, do not imply automatic capture unless the host has its own automation or the user adds one.
• Neo4jStore is experimental. Do not present it as graph-native spreading-activation persistence yet.

Workflow

1. Identify the host: Claude Code, Cursor, Codex, OpenClaw, or generic MCP.
2. If the repo is available locally, read README.md, docs/MCP_INTEGRATION.md, and docs/codex.md before answering.
3. Give the default install path first. Only fall back to manual config snippets if the user cannot use the control plane.
4. Explain the operating loop:
   • mnemos_retrieve at task start
   • mnemos_store for durable facts only
   • mnemos_consolidate before finishing substantial work
   • mnemos_inspect when a stored fact looks wrong
5. Read references/hosts.md for host-specific config snippets and caveats, especially the OpenClaw / ClawHub self-install flow when the agent must bootstrap itself.
6. Read references/operations.md for automation, capture-mode, storage guidance, and troubleshooting.

Avoid

• Do not tell users to manually type memories as the primary workflow.
• Do not recommend SimpleEmbeddingProvider for production retrieval quality.
• Do not upgrade users to Qdrant or Neo4j unless they actually need a non-SQLite profile.