RRememorio Blog

Source notes · Seven-part route

How Hermes Agent runs, learns, and delivers

The series keeps one repair task in view: first follow a request through the model, tools, and state; then see how turn finalization can trigger candidate experience, a new or revised Skill, and finally an adoption decision backed by independent evidence.

Recommended order: Parts I–IV establish the runtime and learning path. Parts V–VII explain data roles, capability revision, and what must be true before a change counts as delivered.

Overview of the Hermes Agent runtime, task entry points, state stores, and background learning

Reading route

From one task to a verifiable self-evolution loop

Each card opens a standalone article. This page only supplies the map, order, and boundaries.

Hermes finalization, persistence, and independent background paths01 · Task runtimeHow one task runs and leaves useful experience behind

Follow one repair through model input, tool execution, history, multiple entry points, and background maintenance.

Read Part I → Instruction sources and durable state responsibilities02 · Instructions and stateWhere should information live?

Separate the system prompt, context, memory, SessionDB, and Skills by role and lifetime.

Read Part II → Task finalization, conditional review, and periodic curation03 · Runtime learningFrom Task Finalization to Background Learning

See how turn traces enter conditional review, separating pending writes from independent periodic curation.

Read Part III → Learning sources, Skill writes, and later discovery04 · Skill creationHow a /learn request becomes a Skill

Trace an explicit learning request through evidence collection, file generation, and reusable capability storage.

Read Part IV → Evaluation sources and three dataset roles05 · Evaluation dataHow train, validation, and holdout differ

Keep one dataset from becoming the exercise, the answer key, and the victory claim at once.

Read Part V → DSPy parameters and the GEPA reflection loop06 · Capability revisionHow DSPy and GEPA rewrite a Skill

Compare compilation, reflection, candidate selection, and the implementation boundary that is actually connected.

Read Part VI → Opt-in evidence recording and bounded turn-end nudges07 · Verification and adoptionWhat counts as delivery after code changes

Inspect project commands, hermes verify, and opt-in evidence recording. Turn-end checks are off by default and provide bounded nudges when enabled; delivery still depends on actual results.

Read Part VII →