Adding a capture language

Auto-capture stores decisions, errors, todos, facts, insights, and preferences. English, Vietnamese, and Chinese each have a CapturePolicy in src/neural_memory/mcp/capture_policy.py. Japanese and Korean are recognized and reported as unsupported. They have no patterns yet.

analyze_text_for_memories asks get_policy(text) for the input floor (20 characters for Latin and Vietnamese, 8 for Chinese). It then runs every registered policy, category by category: decision, error, todo, fact, insight, preference. Inside a category the order is the registry order (en, vi, zh). A sentence that mixes languages keeps a match from each policy. Dedup still happens once, after the scan.

get_policy(text).detect_patterns(text) scans only that language. Use it when a caller already knows the script. The auto-capture entry point does not, because dropping the other policies would change mixed text.

Auto-save triggers in core/trigger_engine.py are a separate detector. A capture policy does not fire them.

Add a language

Subclass CapturePolicy, supply specs(), and register the instance. No edit to auto_capture.py is required.

from neural_memory.mcp.capture_policy import (
    ChinesePolicy,
    register_policy,
)

class JapanesePolicy(ChinesePolicy):
    language_code = "ja"

    def specs(self):
        return _JAPANESE_SPECS  # same six groups, Japanese markers


register_policy(JapanesePolicy())

ChinesePolicy already strips trailing particles and uses the short CJK length floor. A Latin-like language subclasses LatinPolicy instead and inherits the 20-character input floor. Override quality_gate to return 0.0 when a capture should be dropped, or a multiplier such as Vietnamese 0.55. Override empty_hint when an empty result needs a language-specific explanation.

Patterns should require an explicit marker. A descriptive verb in ordinary prose will be stored as a memory at high confidence.

Until a Japanese or Korean policy is registered, detect_language still returns ja or ko, get_policy falls back to Latin for the length floor, and the empty-result hint says the script is not supported.