solohq-memory 0.1.0

Context memory system for AI agents — persistent context graph with episodes, artifacts, relationships, and staleness tracking

solohq-memory

Core context memory library for AI agents. Persistent context graph with episodes, artifacts, relationships, and staleness tracking.

Installation

pip install solohq-memory

# With LLM provider extras
pip install solohq-memory[anthropic]
pip install solohq-memory[openai]
pip install solohq-memory[google]
pip install solohq-memory[all]

Quick Start

from solohq_memory import (
    ContextMemoryManager,
    SqliteVecStorage,
)

# You need to provide LLM and embedding implementations
# that satisfy LLMProvider and EmbeddingProvider protocols

storage = SqliteVecStorage("memory.db", dimensions=1536)
manager = ContextMemoryManager(
    storage=storage,
    llm=your_llm_provider,
    embedder=your_embedding_provider,
)

# Classify a message to find or create a context
result = await manager.classify("Tell me about Python decorators")
context = result.context

# Record an episode
await manager.record_episode(
    context_id=context.id,
    user_message="Tell me about Python decorators",
    assistant_message="Decorators are a way to modify functions...",
)

# Search across all contexts
results = await manager.search("decorators", limit=5)

Key Components

• ContextMemoryManager — main orchestrator for context operations
• SqliteVecStorage — SQLite + sqlite-vec storage with vector search
• LLMClassifier — 3-tier context routing: stable shortcut, embedding similarity, LLM fallback
• ArtifactManager — versioned content objects with rollback support
• EpisodeRecorder — conversation recording
• BoundaryDetector — topic shift detection (heuristics + LLM)
• FileIndexer — file indexing (on-disk references) and upload (full content)

Interfaces

All providers are protocol-based and pluggable:

• StorageProvider — CRUD + vector search
• LLMProvider — complete() and complete_json()
• EmbeddingProvider — embed(), embed_batch(), dimensions
• BlockStorageProvider — session block persistence

License

MIT