G.R.I.L.L.O. — Background Autonomous Agent

G.R.I.L.L.O. is Synth’s background “inner life” subsystem. Instead of only reacting to incoming messages, Grillo periodically generates its own introspective prompts — called beats — that let Synth reflect, consolidate memories, get curious, and keep evolving even when nobody is talking to her.

Each beat is enqueued as a low-priority message and flows through the normal message chain, so it uses Synth’s full context pipeline (persona, emotions, memories, diary). Beats never interrupt a real conversation.

How it works

The Grillo core (grillo_impl.py, wrapped by grillo_plugin.py) runs a scheduler on a fixed interval (GRILLO_BEAT_INTERVAL). On each tick it weighted-selects a beat type and asks the matching beat sub-plugin to build a prompt. Beat sub-plugins are discovered dynamically through the plugin registry — adding a new one is enough to make it available.

Beat types

Each beat sub-plugin lives in its own sub-folder under plugins/grillo/ with a dedicated guide (linked below).

Beat sub-plugin	Beat type	What it does
`grillo_tag`	`tag_elaboration`	Reflect on recent conversation topics/tags.
`grillo_compactor`	`memory_consolidation`	Group old memories by tag and synthesize a compacted memory (nightly).
`grillo_diary_consolidator`	`diary_consolidation`	Merge fragmented diary rows of a past day into one coherent entry.
`grillo_self_reflection`	`self_reflection`	Emotional self check-in, written to the diary.
`grillo_curiosity`	`curiosity`	Explore emergent questions and write a diary entry.
`grillo_relationship`	`relationship`	Reflect on recent interaction patterns.
`grillo_temporal_reflection`	`temporal_reflection`	Notice the time gap since the last user message (ignores routine gaps < 12h).
`grillo_dream`	(daily, ~05:00)	Generate a “dream” diary entry from recent fragments; also a Recon contributor.
`grillo_growth`	(weekly)	Reflect on the week and rewrite Synth’s self-growth reflection and likes/dislikes.
`grillo_chat_observer`	(periodic)	Sample recent chat snippets and propose them for processing.
`grillo_llm_failure_recovery`	(periodic)	Detect recent LLM fallback failures and regenerate the proper reply.

history_evaluator is a non-LLM helper that formats the last N chat messages into a concise reflection prompt used by several beats.

Configuration

Key	Purpose
`GRILLO_BEAT_INTERVAL`	Seconds between scheduler ticks.
`GRILLO_CORTEX`	LLM engine used by Grillo beats.
`GRILLO_ALLOWED_ACTIONS` / `GRILLO_ALLOWED_SECURITY_LEVEL`	Which actions beats may execute.
`GRILLO_COMPACT_ENABLED` / `GRILLO_COMPACT_AGE_DAYS`	Memory-compaction beat toggles.
`GRILLO_OBSERVER_STORE_MEMORIES`	Whether the chat observer stores passive memories.
`GROWTH_MODE`	Self-growth approval mode (`off` / `on` / `request`).

Beats are logged to grillo_activity_log, grillo_beats, and grillo_action_execs.