Memory as a continuous learning substrate

[sghng]

My projects are highly convention-driven and evolving. Practices developed in one project, such as experiment workflows, review methods, writing/R&R lessons, and cluster procedures, often become conventions in others. Agent behavior therefore evolves primarily through shared practices, not merely through accumulating project history.

I see this as related to the Hermes vision of self-reflection loops that improve agent behavior and skills over time. Memory should support continual learning, not only preserve records of past truth.

I prefer one bank, but I want retention and retrieval to be much more targeted and intentional:

• Retention should distinguish project-specific records from lessons deliberately promoted to shared practices.
• Retrieval should let the agent explicitly choose among:
  1. This project: its decisions, progression, key numbers, and results.
  2. Other projects: intentionally search for analogous procedures or compare how similar work was handled elsewhere.
  3. Shared practices: search across the bank for reusable lessons and conventions, without simply retrieving episodic transaction records from whichever projects happen to match semantically.
• Provenance should remain prominent to the agent: source project, document, date, and supporting artifacts. Consolidated observations and reflected answers should not erase that origin. Otherwise, cross-project learning can easily become source confusion and hallucination.

What architecture do you recommend for this?

• How should shared practices be modeled and promoted?
• How should observation_scopes preserve project boundaries while enabling bank-wide practice learning?
• How can the same intentional scopes be applied across recall, reflect, mental models/knowledge pages, and automatic injection?
• How can retrieval results and reflected answers keep provenance visible?
• Which of these capabilities exist in the core API and coding-agent integration today, and where would custom tooling or new first-class retrieval modes be needed?
  Are there docs for this use case? If not, I would welcome guidance or first-class support for project-specific, cross-project analogy, and shared-practice retrieval as distinct modes.