🧠 Agent Memory Store

Give your AI agents long-term memory that actually works.

---

🏗️ Architecture

graph LR
    subgraph "Agent A (Private)"
    A1["Memory A1"]
    A2["Memory A2"]
    A3["..."]
    end
    
    subgraph "Agent B (Private)"
    B1["Memory B1"]
    B2["Memory B2"]
    end
    
    subgraph "Shared Space"
    S1["Shared Memory"]
    S2["Shared Memory"]
    end
    
    A1 --> S1
    B1 --> S1
    A2 -.->|not visible| B1

Loading

Each agent has completely isolated private memory. Shared space is opt-in.

---

Built this because I was tired of agents forgetting everything between conversations. No magic, no vector embeddings — just a simple key-value store with proper isolation between agents.

When you run multiple AI agents, they inevitably step on each other's toes. Agent A reads Agent B's memories, shared context gets mixed up, and suddenly your customer support bot is talking about internal devops stuff.

This library solves that. 🎯

✨ Features

🔒 Agent Isolation	Each agent gets its own private memory space
🤝 Shared Spaces	Optional shared areas for inter-agent communication
💬 Session Tracking	Group memories by conversation
⏳ TTL Support	Auto-expire old memories
⚡ Zero Config	In-memory by default, no external services needed
💾 Pluggable Backends	In-memory, SQLite, JSON file

📦 Install

pip install agent-memory-store

Requires Python 3.9+

🚀 Quick Example

from agent_memory_store import AgentMemoryStore

store = AgentMemoryStore()

# Give an agent its own memory
memory = store.get_agent_memory("support-bot")

# Remember stuff
memory.add("Customer asked about enterprise pricing")
memory.add("FAQ: cancellation takes 24h to process")

# Search later
results = memory.search("pricing")

That's it. No setup, no config, no external services.

---

🔒 Multi-Agent Isolation

Each agent's memory is completely private by default:

store = AgentMemoryStore()

sales = store.get_agent_memory("sales-bot")
support = store.get_agent_memory("support-bot")

sales.add("Q3 target: $500k")
support.add("Refund policy updated")

# They can't see each other's memories
sales.search("refund")   # → []
support.search("sales")  # → []

🤝 Shared Memory

When agents need to share knowledge:

store = AgentMemoryStore()

# Create a shared space
shared = store.get_shared_memory("product-knowledge")
shared.add("🚀 Product launch: May 15th")

# Grant access to multiple agents
sales = store.get_agent_memory("sales-bot", shared_spaces=["product-knowledge"])
support = store.get_agent_memory("support-bot", shared_spaces=["product-knowledge"])

# Both see the shared memory
sales.search("launch")   # ✅ finds it
support.search("launch") # ✅ finds it

💬 Session Tracking

Group memories by conversation:

memory = store.get_agent_memory("support-bot")

with memory.session("conv-12345") as s:
    s.add("Customer: I can't login")
    s.add("Solution: Password reset email sent")

# Pull all memories from a specific conversation
memory.get_by_session("conv-12345")

⏳ Auto-Expiring Memories

from datetime import timedelta

memory.add("Temp context for current task", ttl=timedelta(hours=1))
# Poof. Gone after 1 hour.

---

💾 Backends

Backend	Use Case	Config
memory (default)	Development, testing	AgentMemoryStore()
sqlite	Persistent, single-process	AgentMemoryStore(backend="sqlite", path="./memories.db")
json	Persistent, human-readable	AgentMemoryStore(backend="json", path="./memories.json")

---

📖 API

AgentMemoryStore

store = AgentMemoryStore()

memory = store.get_agent_memory(agent_id, shared_spaces=[...])
shared = store.get_shared_memory(space_id)

store.list_agents()        # → ["sales-bot", "support-bot"]
store.list_shared_spaces() # → ["product-knowledge"]

AgentMemory

memory.add(content, metadata=None, ttl=None, session_id=None)
memory.search(query, limit=10)   # → list[MemorySearchResult]
memory.get_all()                  # → list[Memory]
memory.get_by_session(session_id) # → list[Memory]
memory.delete(memory_id)          # → bool
memory.clear()                    # clear all for this agent

SharedMemory

shared.add(content, metadata=None, ttl=None)
shared.search(query, limit=10)
shared.get_all()
shared.clear()

---

⚠️ What This Is Not

• ❌ Not a vector database — no embeddings, no semantic search
• ❌ Not for huge data — it's an in-memory store with optional persistence
• ❌ Not distributed — single process only (for now)

✅ What This Is

• ✅ Fast — just Python dicts under the hood
• ✅ Simple — 5 minute learning curve
• ✅ Isolated — agents can't accidentally read each other's stuff
• ✅ Zero dependencies — only Python stdlib

---

📄 License

MIT — do whatever you want with it.