agent-memory-toolkit

---

🎯 Why agent-memory-toolkit?

Most agent memory is just "dump everything in a vector DB and pray." That doesn't scale.

agent-memory-toolkit uses hybrid retrieval:

• 🔍 BM25 for exact keyword matches
• 🧬 Vector search for semantic similarity
• 🕸️ Knowledge graph for relational context
• ⚡ RRF fusion to combine results intelligently
• 📉 Ebbinghaus decay so recent memories surface naturally

95.2% R@5 on LongMemEval-S — state-of-the-art recall for long-term agent memory

---

🔐 Local-First. Your Data Stays Yours.

No cloud. No API calls for storage. Everything runs on SQLite.

• ✅ Works offline
• ✅ GDPR-friendly
• ✅ Airgapped environments
• ✅ Full control over your data

---

✨ Features

---

📦 Install

pip install agent-memory-toolkit

With all features:

pip install agent-memory-toolkit[all]

---

🚀 Quick Start

from agent_memory_toolkit import MemoryStore

# Create a local memory store
store = MemoryStore("memories.db", auto_embed=True)

# Add memories
store.add("User prefers dark mode and vim keybindings")
store.add("Project deadline is Friday, client is Acme Corp")
store.add("Last meeting discussed Q4 roadmap")

# Hybrid search (BM25 + vectors + recency decay)
results = store.search("vim preferences", mode="hybrid")

for r in results:
    print(f"[{r.score:.2f}] {r.memory.content}")

Extract Structured Memories

from agent_memory_toolkit import MemoryExtractor

extractor = MemoryExtractor()

text = """
Hi, I'm Sarah Chen. I work as a Senior Engineer at TechCorp.
I prefer Python over JavaScript and usually work 9-5 PST.
"""

memories = extractor.extract(text)

for m in memories.memories:
    print(f"[{m.domain.value}] {m.key}: {m.value}")
# [biography] name: Sarah Chen
# [work] role: Senior Engineer
# [preferences] preferred_language: Python

---

🏗️ Architecture

┌─────────────────────────────────────────────────────────────────────────────┐
│                             AGENTMEMORY                                     │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│  ┌─────────────┐   ┌──────────────┐   ┌───────────────┐   ┌─────────────┐  │
│  │ Extraction  │   │   Storage    │   │  Compression  │   │  Security   │  │
│  │   Module    │   │    Store     │   │    Engine     │   │   Guard     │  │
│  │             │   │              │   │               │   │             │  │
│  │ • Rule-based│   │ • SQLite     │   │ • Token aware │   │ • Poison    │  │
│  │ • LLM-based │   │ • FTS5/BM25  │   │ • Importance  │   │   detection │  │
│  │ • Hybrid    │   │ • Vectors    │   │   ranking     │   │ • Confidence│  │
│  │ • 6 domains │   │ • RRF Fusion │   │ • Strategies  │   │   scoring   │  │
│  └─────────────┘   └──────────────┘   └───────────────┘   └─────────────┘  │
│         │                  │                   │                  │        │
│         └──────────────────┴───────────────────┴──────────────────┘        │
│                                    │                                        │
│                        ┌───────────┴───────────┐                           │
│                        │   Team Memory Store   │                           │
│                        │                       │                           │
│                        │ • Git-like branching  │                           │
│                        │ • Conflict resolution │                           │
│                        │ • Filesystem sync     │                           │
│                        │ • Access control      │                           │
│                        └───────────────────────┘                           │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘

---

📊 Benchmarks

Hybrid retrieval with RRF fusion significantly outperforms single-strategy approaches.

---

⚡ Performance

---

📖 API Reference

MemoryStore

from agent_memory_toolkit import MemoryStore

store = MemoryStore(
    db_path="memories.db",
    auto_embed=True,
    embedding_model="all-MiniLM-L6-v2"
)

# Core operations
store.add(content, metadata=None)
store.get(memory_id)
store.update(memory_id, content=None, metadata=None)
store.delete(memory_id)

# Search modes
store.search(query, mode="hybrid")  # BM25 + vectors + decay
store.search_fts(query)             # BM25 only
store.search_vector(query)          # Vectors only

MemoryExtractor

from agent_memory_toolkit import MemoryExtractor, CognitiveDomain

extractor = MemoryExtractor(mode="rule")  # or "llm", "hybrid"
result = extractor.extract(text)

MemoryGuard

from agent_memory_toolkit import MemoryGuard, SecurityLevel

guard = MemoryGuard(level=SecurityLevel.HIGH)
result = guard.validate_content(content)

if result.is_safe:
    store.add(content)

TeamMemoryStore

from agent_memory_toolkit.team import TeamMemoryStore

store = TeamMemoryStore("team.db", agent_id="alice")

# Git-like operations
store.create_branch("experiment")
store.checkout("experiment")
store.commit("Added new findings")
store.push("/shared/memories")
store.pull("/shared/memories")

---

📂 Examples

See examples/ for working demos:

• basic_usage.py — Getting started
• team_collaboration.py — Multi-agent workflows
• secure_memory.py — Security validation
• compress_context.py — Context compression

---

🧪 Testing

pytest
pytest --cov=agent_memory_toolkit

---

🤝 Contributing

1. Fork the repo
2. Create a feature branch (git checkout -b feature/amazing)
3. Commit changes (git commit -m 'Add amazing feature')
4. Push to branch (git push origin feature/amazing)
5. Open a Pull Request

---

📄 License

MIT License — see LICENSE

---