hermes - 💡(How to fix) Fix [Feature]: Inject mental model into context when using Hindsight

Problem or Use Case

When using hindsight, hermes uses the user's prompt with the recall and reflect endpoints. However, hindsight mental models are powerful and could provide valuable context.

Proposed Solution

Looking into this, honcho uses static text (Who is this person? What are their preferences, goals, and working style?) to inject information about the user into context. I have used this same question in hindsight to create a mental model with very good results. The proposed solution would allow user's to specify an arbitrary mental model of their own creation and choosing to inject into context, similar to honcho.

The proposed solution includes:

• Create a new hindsight environmental variable (HINDSIGHT_API_INJECTION_MODEL_ID) allowing a user to specify the ID of the mental model
• Update the hindsight provider. If HINDSIGHT_API_INJECTION_MODEL_ID exists, fetch and cache the mental model and to inject the returned mental model into context.
• Keep the existing recall / reflect behavior. This acts as both a fallback behavior and can continue to supplement the injected context.
• Add/update guards to ensure that injected context does not exceed max characters.
• Update documentation with configuration and sample mental model creation.

Alternatives Considered

I experimented with using the same question against the reflect endpoint. The results were similar in quality to the mental model, but was signfiicantly slower than fetching the mental model.

Feature Type

Configuration option

Scope

Medium (few files, < 300 lines)

Contribution

• I'd like to implement this myself and submit a PR

Debug Report (optional)