OpenMem: Persistent Neuro-Symbolic Memory

A high-performance, production-ready memory coprocessor for LLM agents.

OpenMem (PNME) provides a durable long-term memory layer that survives across sessions, crashes, and restarts. By combining Hyperdimensional Computing (HDC) with symbolic triples, it offers high-speed retrieval, principled association, and structured fact management.

Installation

# Clone the repository
git clone https://github.com/OpenMem-Project/OpenMem
cd OpenMem

# Install dependencies and package
pip install -e .

Core Features

• Principled HDC Encoding: Uses 10,000-dimensional bipolar vectors with deterministic hash-based seeds and role-vector binding.
• Hybrid Retrieval: Combines exact symbolic filtering with sub-symbolic HDC unbinding and multi-factor ranking (recency, strength, provenance).
• Audit & Analytics: Integrated access logging and event tracking for memory lifecycle analysis.
• Portability: JSONL export/import support for cross-system memory migration.
• Safety: Automated scrubbing of sensitive information (secrets, keys) before storage.

Agent Integrations

1. Claude Code

OpenMem integrates with Claude via a structured tool adapter. To use it, import the adapter and register the tools:

from pnme.integrations.claude_tools import ClaudeMemoryAdapter

# Initialize adapter
adapter = ClaudeMemoryAdapter(db_path="my_memory.db")
tools = adapter.get_tool_definitions()

# In your Claude tool handler:
def on_tool_call(name, args):
    return adapter.handle_tool_call(name, args)

Available Tools:

• memory_store: Save a specific fact (S, R, O).
• memory_absorb: Bulk-extract facts from a text block.
• memory_query: Pattern-based retrieval.
• memory_hydrate: Inject context into the prompt.

2. OpenClaw

Register OpenMem as a plugin in your OpenClaw setup:

from pnme.integrations.openclaw_plugin import setup_plugin

# Setup the memory plugin
memory_plugin = setup_plugin({"db_path": "openclaw_mem.db"})
agent.register_plugin(memory_plugin)

Quick Start (Python)

from pnme.api import PNME

memory = PNME()

# Store a fact
memory.store("DeepSeek-V3", "released_by", "DeepSeek")

# Absorb facts from text
memory.absorb("Claude 3.5 Sonnet is a model by Anthropic. It supports computer use.")

# Retrieve hydrated context
prompt = "Tell me about Anthropic models."
hydrated_prompt = memory.hydrate(prompt)

License

MIT