ontomem 0.1.0

A self-consolidating memory layer for AI agents with schema-first design, intelligent merging, and hybrid search capabilities

🧠 Ontomem: The Self-Consolidating Memory Layer

中文版本 | English

Give your AI agent a "coherent" memory, not just "fragmented" retrieval.

Traditional RAG (Retrieval-Augmented Generation) systems retrieve text fragments. Ontomem maintains structured entities using Pydantic schemas and intelligent merging algorithms. It automatically consolidates fragmented observations into complete knowledge graph nodes.

It doesn't just store data—it continuously "digests" and "organizes" it.

---

✨ Why Ontomem?

🧩 Schema-First & Type-Safe

Built on Pydantic. All memories are strongly-typed objects. Say goodbye to {"unknown": "dict"} hell and embrace IDE autocomplete and type checking.

🔄 Auto-Consolidation

When you insert different pieces of information about the same entity (same ID) multiple times, Ontomem doesn't create duplicates. It intelligently merges them into a Golden Record using configurable strategies (field overrides, list merging, or LLM-powered intelligent fusion).

🔍 Hybrid Search

• Key-Value Lookup: O(1) exact entity access
• Vector Search: Built-in FAISS indexing for semantic similarity search, automatically synced

💾 Stateful & Persistent

Save your complete memory state (structured data + vector indices) to disk and restore it in seconds on next startup.

---

🚀 Quick Start: Building a "Self-Improving" Experience Library

Imagine an AI coding agent that debugs issues. Without memory, it repeats the same trial-and-error process every time. With Ontomem, it builds a persistent "Debugging Playbook" that evolves with each new problem encountered.

1. Define Your Experience Schema

from pydantic import BaseModel
from typing import List, Optional

class BugFixExperience(BaseModel):
    """A living record of debugging knowledge."""
    error_signature: str            # Key: e.g., "ModuleNotFoundError: pandas"
    root_causes: List[str]          # Different reasons this error can occur
    solutions: List[str]            # Multiple working solutions discovered
    prevention_tips: str            # Synthesized understanding of how to avoid it
    last_updated: Optional[str] = None

2. Initialize with LLM-Powered Merging

We use the LLM.BALANCED strategy so Ontomem doesn't just list solutions—it synthesizes them into coherent, actionable guidance.

from ontomem import OMem, MergeStrategy
from langchain_openai import ChatOpenAI, OpenAIEmbeddings

experience_memory = OMem(
    memory_schema=BugFixExperience,
    key_extractor=lambda x: x.error_signature,
    llm_client=ChatOpenAI(model="gpt-4o"),
    embedder=OpenAIEmbeddings(),
    merge_strategy=MergeStrategy.LLM.BALANCED
)

3. The Agent Learns Over Time

Day 1: The First Encounter

The agent encounters ModuleNotFoundError for pandas and fixes it with pip install.

# Experience 1: Initial observation
experience_memory.add(BugFixExperience(
    error_signature="ModuleNotFoundError: No module named 'pandas'",
    root_causes=["Missing library in environment"],
    solutions=["Run: pip install pandas"],
    prevention_tips="Always check requirements.txt before running code."
))

Day 2: New Context, Different Fix

The agent encounters the same error in a Docker container where pip fails, but apt-get install python3-pandas works.

# Experience 2: Different context, same error
experience_memory.add(BugFixExperience(
    error_signature="ModuleNotFoundError: No module named 'pandas'",
    root_causes=["Package not in system Python", "Binary incompatibility with pip"],
    solutions=["Run: apt-get install python3-pandas", "Use system package manager in containers"],
    prevention_tips="In containerized environments, prefer system packages for compiled dependencies."
))

Day 3: Agent Seeks Wisdom

When a new agent instance encounters the same error, it queries the evolved knowledge base:

# Retrieve consolidated wisdom
guidance = experience_memory.get("ModuleNotFoundError: No module named 'pandas'")

print("Root Causes:")
for cause in guidance.root_causes:
    print(f"  - {cause}")
# Output:
#   - Missing library in environment
#   - Package not in system Python
#   - Binary incompatibility with pip

print("\nSolutions:")
for i, solution in enumerate(guidance.solutions, 1):
    print(f"  {i}. {solution}")
# Output:
#   1. Run: pip install pandas (standard approach)
#   2. Run: apt-get install python3-pandas (for system Python)
#   3. Use system package manager in containers

print("\nPrevention Tips:")
print(guidance.prevention_tips)
# Output: "Check requirements.txt before running code. 
#         In containers, prefer system packages for compiled dependencies.
#         Consider using virtual environments to isolate dependencies."

Day 4: Semantic Search for Similar Problems

The agent doesn't remember the exact error, but can search by concept:

# Semantic search: Find solutions for import-related issues
similar_issues = experience_memory.search(
    "Python module import failures dependency missing",
    k=5
)

print(f"Found {len(similar_issues)} related debugging experiences")

The agent went from "trial and error" to "informed decision-making". No boilerplate. No manual consolidation. Just add experiences and let Ontomem synthesize wisdom.

---

🔍 Semantic Search

Build an index and search by natural language:

# Build vector index
memory.build_index()

# Semantic search
results = memory.search("Find researchers working on transformer models and attention mechanisms")

for researcher in results:
    print(f"- {researcher.name}: {researcher.research_interests}")

---

🛠️ Merge Strategies

Choose how to handle conflicts:

Strategy	Behavior	Use Case
FIELD_MERGE	Non-null overwrites, lists append	Simple attribute collection
KEEP_NEW	Latest data wins	Status updates (current role, last seen)
KEEP_OLD	First observation stays	Historical records (first publication year)
LLM.BALANCED	LLM-driven semantic merging	Complex synthesis, contradiction resolution

# Example: LLM intelligently merges conflicting bios
memory = OMem(
    ...,
    merge_strategy=MergeStrategy.LLM.BALANCED
)

---

💾 Save & Load

Snapshot your entire memory state:

# Save (structured data → memory.json, vectors → FAISS indices)
memory.dump("./researcher_knowledge")

# Later, restore instantly
new_memory = OMem(...)
new_memory.load("./researcher_knowledge")

---

📊 Ontomem vs Traditional Approaches

Feature	Traditional Vector DB	Ontomem 🧠
Storage Unit	Text chunks	Structured Objects
Deduplication	Manual or via embeddings	Native, ID-based
Updates	Append-only (creates dupes)	Auto-merge (upsert)
Query Results	Similar text fragments	Complete entities
Type Safety	❌ None	✅ Pydantic
Indexing	Manual sync needed	✅ Auto-synced

---

🎯 Use Cases

🤖 AI Research Assistant

Consolidate researcher profiles, papers, and citations from multiple sources.

👤 Personal Knowledge Graph

Build a living profile of contacts, their preferences, skills, and interaction history from conversations.

🏢 Enterprise Data Hub

Unify customer/employee records from CRM, email, support tickets, and social media.

🧠 AI Agent Long-Term Memory

An autonomous agent accumulates experiences and observations—Ontomem keeps them organized and searchable.

---

🔧 Installation

pip install ontomem

Or with uv:

uv add ontomem

Requirements:

• Python 3.11+
• LangChain (for LLM integration)
• Pydantic (for schema definition)
• FAISS (for vector search)

---

📚 API Reference

Core Methods

add(items: Union[T, List[T]]) → None

Add item(s) to memory. Automatically merges duplicates by key.

memory.add(ResearcherProfile(...))
memory.add([item1, item2, item3])

get(key: Any) → Optional[T]

Retrieve an entity by its unique key.

researcher = memory.get("yann_lecun_001")

build_index(force: bool = False) → None

Build or rebuild the vector index for semantic search.

memory.build_index()  # Build if clean
memory.build_index(force=True)  # Force rebuild

search(query: str, k: int = 5) → List[T]

Semantic search over all entities.

results = memory.search("transformers and attention", k=10)

dump(folder_path: Union[str, Path]) → None

Save memory state (data + index) to disk.

memory.dump("./my_memory")

load(folder_path: Union[str, Path]) → None

Load memory state from disk.

memory.load("./my_memory")

remove(key: Any) → bool

Remove an entity by key.

success = memory.remove("yann_lecun_001")

clear() → None

Clear all entities and indices.

memory.clear()

Properties

keys: List[Any]

All unique keys in memory.

items: List[T]

All entity instances.

size: int

Number of entities.

---

🤝 Contributing

We're building the next generation of AI memory standards. PRs and issues welcome!

---

📝 License

MIT License - See LICENSE file for details.

---

Built with ❤️ for AI developers who believe memory is more than just search.