Summary
everalgo/user_memory/prompts/zh/reflect.py re-exports the English REFLECT_EPISODE_PROMPT / REFLECT_EPISODE_UPDATE_PROMPT verbatim, per its own docstring:
"""Chinese prompts for EpisodeReflector — re-exports English (language-neutral for now)."""
Unlike its sibling zh/* modules (profile.py, episode.py, atomic_fact.py, the parser prompts, etc.), which all ship dedicated Chinese-localized prompt text, reflect.py has no zh-specific instructions at all — the English prompt is used unmodified for Chinese-language input.
Observed impact
Running EpisodeReflector against 100%-Chinese episode input, the merged narrative it produces intermittently mixes English phrasing into an otherwise Chinese narrative — a single paraphrased sentence (e.g. a decision/status summary) can render half in Chinese and half in English within the same output, even though every source episode was pure Chinese. This is consistent with the English-only instruction text leaking style/language into generation, since neither REFLECT_EPISODE_PROMPT nor REFLECT_EPISODE_UPDATE_PROMPT contains any "respond in the input language" instruction.
Downstream, this makes reflected narratives unreliable to consume as a clean-language data source for CJK deployments (we ended up excluding this field from a downstream pipeline because of the mixing).
Comparison with the other zh re-export
everalgo/clustering/prompts/zh/cluster.py is the only other module in the tree that re-exports the English template, but it documents why that's fine:
"""Chinese clustering prompts — re-export of the English template.
The clustering LLM template is language-neutral: responses match the candidate corpus' language...
"""
reflect.py's "for now" phrasing reads as an acknowledged gap rather than the same kind of deliberate, verified design decision — which matches what we're observing empirically (unlike clustering output, reflect output does not reliably match the input language).
Suggested fix
Doesn't require a full Chinese translation of the prompt — a single added instruction to the shared English template (e.g. "Write the narrative in the same language as the input episodes.") should close the gap while preserving the existing language-neutral prompt design.
Environment
everalgo installed as a dependency of EverOS/evermem (pip show everalgo via the evermem venv)- Affected:
everalgo.user_memory.reflect.EpisodeReflector (both full-merge and incremental-update code paths)
Summary
everalgo/user_memory/prompts/zh/reflect.pyre-exports the EnglishREFLECT_EPISODE_PROMPT/REFLECT_EPISODE_UPDATE_PROMPTverbatim, per its own docstring:"""Chinese prompts for EpisodeReflector — re-exports English (language-neutral for now)."""Unlike its sibling
zh/*modules (profile.py,episode.py,atomic_fact.py, the parser prompts, etc.), which all ship dedicated Chinese-localized prompt text,reflect.pyhas no zh-specific instructions at all — the English prompt is used unmodified for Chinese-language input.Observed impact
Running
EpisodeReflectoragainst 100%-Chinese episode input, the merged narrative it produces intermittently mixes English phrasing into an otherwise Chinese narrative — a single paraphrased sentence (e.g. a decision/status summary) can render half in Chinese and half in English within the same output, even though every source episode was pure Chinese. This is consistent with the English-only instruction text leaking style/language into generation, since neitherREFLECT_EPISODE_PROMPTnorREFLECT_EPISODE_UPDATE_PROMPTcontains any "respond in the input language" instruction.Downstream, this makes reflected narratives unreliable to consume as a clean-language data source for CJK deployments (we ended up excluding this field from a downstream pipeline because of the mixing).
Comparison with the other zh re-export
everalgo/clustering/prompts/zh/cluster.pyis the only other module in the tree that re-exports the English template, but it documents why that's fine:reflect.py's "for now" phrasing reads as an acknowledged gap rather than the same kind of deliberate, verified design decision — which matches what we're observing empirically (unlike clustering output, reflect output does not reliably match the input language).Suggested fix
Doesn't require a full Chinese translation of the prompt — a single added instruction to the shared English template (e.g. "Write the narrative in the same language as the input episodes.") should close the gap while preserving the existing language-neutral prompt design.
Environment
everalgoinstalled as a dependency of EverOS/evermem (pip show everalgovia the evermem venv)everalgo.user_memory.reflect.EpisodeReflector(both full-merge and incremental-update code paths)