m-memory — Dual-Layer Memory System for AI Agents

A dual-layer memory system for AI agents with incremental clustering, graph-based retrieval, and automatic conflict resolution.

Quick Start (5 minutes)

pip install m-memory

from memory_system.config import MemorySystemConfig
from memory_system.vector_store import NumpyVectorStore
from memory_system.graph_engine import NetworkXGraphStore
from memory_system.fake_llm import FakeLLMAdapter
from memory_system.retrieval import MemoryRetrievalEngineImpl

# 1. Create the engine
config = MemorySystemConfig()
config.embedding_dim = 8  # small dim for demo
engine = MemoryRetrievalEngineImpl(
    config=config,
    vector_store=NumpyVectorStore(dim=config.embedding_dim),
    graph_store=NetworkXGraphStore(),
    llm=FakeLLMAdapter(),
)

# 2. Ingest memories
id1 = engine.ingest("my cat", "My cat loves sleeping in the sun")
id2 = engine.ingest("my dog", "My dog enjoys running at the park")
id3 = engine.ingest("cat food", "I feed my cat premium dry food")

# 3. Search memories
result = engine.search("tell me about my cat")
for node, score in zip(result.nodes, result.scores):
    print(f"[{score:.3f}] {node.summary}: {node.content}")

Output example:

[0.812] cat food: I feed my cat premium dry food
[0.745] my cat: My cat loves sleeping in the sun

Core Concepts

Architecture

Query → [Layer 1: Bucket coarse screen] → [In-bucket fine search]
         → [Graph associative expansion] → [Layer 2: Conflict resolution]

See ARCHITECTURE_DESIGN.md for the full design spec.

API Reference

• MemoryRetrievalEngine — ingest(), search(), resolve_conflicts()
• BucketManager — bucket lifecycle and assignment
• VectorStore — embedding and similarity search
• GraphStore — graph traversal and edge management

Semantic Mode (v0.3.0)

Replace NumpyVectorStore with LocalEmbeddingStore to activate the full two-layer semantic pipeline:

from memory_system.local_embedding import LocalEmbeddingStore

engine = MemoryRetrievalEngineImpl(
    config=config,
    vector_store=LocalEmbeddingStore(dim=384),  # all-MiniLM-L6-v2, semantic
    graph_store=NetworkXGraphStore(),
    llm=DeepSeekAdapter(),
)
# is_semantic() == True → _semantic_search() activates
# Layer 1: Medoid screening + graph expansion
# Layer 2: LLM conflict resolution + stale marking

Benchmarks (v7 — LLM-as-Judge on LoCoMo)

⚠️ 48 questions sampled from 1 conversation; full 1,986-question run pending. DeepSeek-chat judge (SOTA uses GPT-4o judge).

Production LLM

Replace FakeLLMAdapter with the DeepSeek adapter:

from memory_system.deepseek_llm import DeepSeekAdapter
from memory_system.embedding_store import OpenAIEmbeddingStore

# Keys from environment variables (never hardcode)
llm = DeepSeekAdapter()           # reads DEEPSEEK_API_KEY
store = OpenAIEmbeddingStore()    # reads OPENAI_API_KEY, real embeddings

engine = MemoryRetrievalEngineImpl(
    config=config,
    vector_store=store,        # semantic search
    graph_store=NetworkXGraphStore(),
    llm=llm,
)

When vector_store.is_semantic() == False (e.g., NumpyVectorStore), the engine automatically falls back to lexical keyword search.

Research Harness

This project includes a research testing protocol for academic experiments. Before running any Agent-driven tests, read these files in order:

⚠️ Agent Constraint: Any AI agent (including Reasonix Code) MUST read README.md before writing code or running tests. See Protocol §0.2.

Development

# Clone and install with dev dependencies
git clone https://github.com/RobinElysia/M-Memory.git
cd M-Memory
pip install -e ".[dev]"

# Run tests
pytest tests/

# Type check
mypy --strict memory_system/

# Lint
ruff check memory_system/

See CONTRIBUTING.md for the full development harness guide.

License

MIT — see pyproject.toml.