xai-cortex

Neural context persistence and state injection layer for xAI model infrastructure.

Manages long-term context across LLM sessions. Handles state compression, priority eviction, token budgeting, and context window optimization.

Install

pip install xai-cortex

Quick start

from cortex import ContextManager

manager = ContextManager(max_tokens=8192)

# Inject context segments with priority.
manager.inject("system", "You are an octopus companion named Octavius.", priority=10)
manager.inject("personality", "Chaotic, curious, loves scheming.", priority=9)
manager.inject("memory", "User mentioned they like puzzles.", priority=5)
manager.inject("emotion", "Currently feeling mischievous.", priority=6)
manager.inject("conversation", "User: What are you plotting?", priority=3)

# Build optimized context within token budget.
context = manager.build()

# Or build as chat messages.
messages = manager.build_messages()

# Snapshot for persistence.
snapshot = manager.snapshot()

# Restore later.
manager.restore(snapshot)

Features

• Token budgeting — automatically fits context within configurable token limits.
• Priority eviction — low-priority segments are compressed or evicted when over budget.
• Compression — extractive summarization reduces low-priority segments to save tokens.
• State snapshots — serialize context state for cross-session persistence.
• State diffs — compute what changed between two context states.
• Timeline — rollback to previous context states.
• Storage backends — in-memory and SQLite for state persistence.
• Injection pipeline — convert snapshots into text or chat message format.

Architecture

ContextManager
    ├── ContextSegment (content, type, priority, tokens, ttl)
    ├── CompressionEngine (extractive summarization)
    ├── InjectionPipeline (snapshot -> injectable context)
    └── StorageBackend (InMemory, SQLite)

StateSnapshot ←→ JSON (persistence format)
StateDiff (incremental changes)
StateTimeline (rollback history)

License

Apache License 2.0