Using with non Hermes harnesses

[kjhughes097]

Is this only for Hermes Agent, or do you see it being used for other harnesses too?
Where would the integration points be for other harnesses, maybe a doc or readme explaining that would help me get started (assuming you see/want it used for other harnesses).

I have my own harness that I'd like to be able to use it with...

[ClaudioDrews]

It's currently Hermes-native — the integration points (Icarus plugin hooks, fabric tools, pre_llm_call injection, fact_store/SQLite schema) are designed around Hermes Agent's plugin architecture.

That said, decoupling the core memory infrastructure from the Hermes-specific integration layer is on the roadmap. The Docker stack (Qdrant + Redis + ARQ worker) is already harness-agnostic — it's the injection hooks that are tightly coupled. With v0.2.0 shipping a reproducible test environment, we can now actually test against alternative harnesses.

I'll post an update here once there's a documented integration path. If you're willing to share details about your harness (language, hook model, how it injects context), that would help prioritize which integration patterns to support first.

[ClaudioDrews]

Memory OS is designed to be harness-agnostic at the Icarus layer. The integration points are:

Inbound (memory → agent): Call pre_llm_call(user_message) before each LLM turn. It queries 4 sources in parallel (Fabric, Qdrant, Sessions, Facts) and returns context to inject into the system prompt. See layers/04-icarus-fabric.md for the full flow.

Outbound (agent → memory): Call post_llm_call(user_msg, assistant_msg) to capture decisions, and on_session_end() for session archival.

The hooks are in icarus/hooks.py — they're plain Python functions with no Hermes-specific dependencies beyond the Icarus plugin convention. Any harness that can call a Python function in the agent lifecycle can integrate.

Would a dedicated docs/harness-integration.md guide be useful? If so, let me know which harness you're using and I'll write one.