Memory Bank supplies eligible cross-job facts to bounded cognitive stages. Firestore remains the provenance anchor and workflow authority.
How Harmonia uses it
- Retrieval is scoped to the exact workspace and brand.
- Returned facts include stable fact IDs and Firestore evidence references.
- Nimi and Ryan may use eligible facts to qualify analysis and strategy recommendations.
- Learning writes admit only typed, evidence-linked decisions and verified outcomes.
- Retrieval and generation are bounded to small result sets.
Raw prompts, transcripts, drafts, logs, replay fixtures, unverified claims, credentials, and private reasoning are excluded from durable memory.
Security and authority
A memory fact cannot approve a strategy, authorize a channel, override policy, change a budget, schedule content, or permit publishing. Outputs that present memory as authorization fail validation.
Failure behavior
Unavailable memory is visible but does not create fictional context. Stages continue only where the contract permits missing advisory memory; they never replace it with recalled facts.
Evidence status
Scope, eligibility, provenance, and failure contracts have automated local coverage. Authenticated retrieval or generation requires captured Memory Bank operation evidence.
Relevant files
agent/harmonia_agent/memory_bank.pyagent/harmonia_agent/model_governance.pyagent/harmonia_agent/stages.pydocs/session-memory.mdxdocs/state-ownership.mdx
Last modified on August 28, 2026
