jyapayne/mnemosyne
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix: scale FM window_size to match real API pressure for fidelity degradation

Browse source 
The FidelityManager's internal pressure calculation uses its own tracked
object tokens divided by window_size, which is always tiny compared to
the real context. Temporarily scale window_size so the FM's pressure
matches the actual API input_tokens/window ratio, triggering L0→L1→L2
degradations when context exceeds 50%.
This commit is contained in:
Joey Yakimowich-Payne 2026-03-13 21:13:38 -06:00
commit 65e4e38a98
1 changed files with 15 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

22
src/mnemosyne/gateway.py
View file

@ -2122,8 +2122,8 @@ def create_app(
"""Update FidelityManager window understanding from actual API usage. """Update FidelityManager window understanding from actual API usage.
After receiving the response, we know the real input token count. After receiving the response, we know the real input token count.
Update the fidelity manager's window_size understanding and schedule Scale the FM's window_size so its internal pressure calculation
degradation if pressure is above NORMAL for the next turn. reflects the real API token usage, then trigger degradation.
""" """
input_tokens = usage.get("input_tokens", 0) input_tokens = usage.get("input_tokens", 0)
if input_tokens <= 0: if input_tokens <= 0:
@ -2132,14 +2132,22 @@ def create_app(
fm = session.fidelity_manager fm = session.fidelity_manager
turn = session.token_state.get("turn", 0) turn = session.token_state.get("turn", 0)
# The FidelityManager tracks its own token budget via registered objects.
# Here we use the real API token count to check if we need proactive degradation.
# If real usage exceeds the fidelity window threshold, trigger degradation now
# so the NEXT turn benefits from reduced content.
pressure_ratio = input_tokens / fm.window_size if fm.window_size > 0 else 1.0 pressure_ratio = input_tokens / fm.window_size if fm.window_size > 0 else 1.0
if pressure_ratio >= fm.threshold_caution: if pressure_ratio >= fm.threshold_caution:
transitions = fm.degrade(turn) # The FM's internal pressure uses total_tokens()/window_size, but
# total_tokens() only counts registered objects (a fraction of the
# real context). Temporarily scale window_size down so the FM's
# pressure matches the real API pressure, then restore it.
obj_tokens = fm.total_tokens()
if obj_tokens > 0:
# Set window_size so obj_tokens/window_size == pressure_ratio
saved_ws = fm.window_size
fm.window_size = max(1, int(obj_tokens / pressure_ratio))
transitions = fm.degrade(turn)
fm.window_size = saved_ws
else:
transitions = fm.degrade(turn)
if transitions: if transitions:
zone = fm.current_pressure() zone = fm.current_pressure()
for _obj_id, old_level, new_level in transitions: for _obj_id, old_level, new_level in transitions: