affective-fly host¶

emotional_memory is the host. affective-fly is the in-process affect source: a reduced Drosophila mushroom-body circuit that returns valence, arousal, and approach/avoid.

This is the inverse of python -m affective_fly run / study, which construct EmotionalMemory inside the fly package. Do not invert that. The fly's Policy and LaunchGate stay at library defaults and are not the product. Phase 6 hypothesis mood taus (300 / 60 / 180 s) are not retuned here.

The contract is HostFrame v1.0 plus host-owned mood_dt, documented in affective-fly's docs/HOST_INTEGRATION.md and examples/host_owns_loop.py.

Installation¶

affective-fly is not on PyPI (and listing it as a locked extra would cycle: fly already depends on this package).

make install-fly
# or:
uv pip install "affective-fly @ git+https://github.com/gianlucamazza/affective-fly.git"

Ownership¶

This package constructs EmotionalMemory and the store.
This package owns time and computes mood_dt (wall-clock or HostFrame timestamps).
Each tick calls the fly (AffectiveLoop / HostAdapter / circuit) and reads valence / arousal / approach-avoid.

FlyAffectHost is the thin wrapper. Pass your own store, embedder, and memory when you already have an engine; pass the same store and embedder instances used to construct that memory (AffectiveLoop cannot read engine internals).

Live tick¶

from affective_fly import HostFrame
from emotional_memory.integrations import FlyAffectHost

host = FlyAffectHost()
frame = HostFrame(
    visual_hash="note-1",  # chosen at creation; required for exact visual replay
    context={
        "context": "journal",
        "note_id": "n1",
        "query": "successful replication",
        "sentiment": 0.8,
    },
)
affect = host.tick_wall_clock(frame, now=10.0)
print(affect.valence, affect.arousal, affect.approach, affect.mood_dt)

now is seconds from time.monotonic() (or a test double). The first tick is mood_dt=0.0 — there is no previous sample to invent.

Replay¶

Replay recorded HostFrame JSON with mood_dt from timestamps. Report an outcome (reward / outcome / pnl) only when you have one — do not invent rewards.

from affective_fly import HostAdapter, HostFrame
from emotional_memory.integrations import FlyAffectHost

frames = [
    HostFrame(
        timestamp="2026-09-12T10:00:00Z",
        visual_hash="exp-042-journal",
        context={
            "context": "journal",
            "note_id": "exp-042-replication",
            "query": "successful replication of experiment 042",
            "sentiment": 1.0,
        },
    ),
    HostFrame(
        timestamp="2026-09-12T14:30:00Z",
        visual_hash="exp-042-review",
        context={
            "context": "review",
            "note_id": "exp-042-replication",
            "query": "replication failed validation",
            "sentiment": -0.9,
            "outcome": -0.8,
        },
    ),
]
host = FlyAffectHost()
affects = host.replay(frames)  # mood_dt: 0.0, then 16200.0

You can also load a JSONL journal with HostAdapter.load_journal(...).

Phase 6 measurement log¶

This host writes fly measure.jsonl when you pass measure_path. The host owns time (mood_dt from wall-clock or HostFrame timestamps); the fly only supplies affect. Hypothesis taus (300 / 60 / 180) are not retuned here.

host = FlyAffectHost(measure_path="measure.jsonl")
host.replay(frames)

python examples/fly_affect_source.py --measure measure.jsonl
python examples/fly_phase6_session.py --measure measure.jsonl
python -m affective_fly calibrate measure.jsonl

fly_phase6_session.py replays HostFrames spanning tens of minutes of mood time (timestamp deltas). A few 0.03 s live ticks is not a Phase 6 session. Do not commit fitted taus or Policy / LaunchGate thresholds from that log.

Defaults that stay frozen¶

Knob	Value	Notes
Fly `MoodField` taus	300 / 60 / 180 s	Phase 6 hypothesis; not lab 8 / 4 / 5
`Policy.threshold_act`	0.2	Unchanged
`Policy.threshold_calm`	0.0	Unchanged
`LaunchGate.threshold_approach`	0.2	Unchanged
Default circuit	`LIFCircuit(n_kc=200, ...)`	Pass `MockFlyCircuit` only in unit tests

This is not a UI or a token-launch product. Circuit CoreAffect is the fast path; Scherer appraisal stays tag-only. Do not pass appraisal= to encode().

Run the example¶

make install-fly
python examples/fly_affect_source.py
python examples/fly_phase6_session.py --measure measure.jsonl
uv run python -m pytest tests/test_fly_adapter.py