AgentDNA

Inspiration

AI agents are becoming increasingly capable, but most agents still suffer from a fundamental limitation: they forget. Conversations, decisions, user preferences, previous actions, and lessons learned are often lost between sessions.

We wanted to build an agent that doesn't simply respond—it remembers, learns from history, and uses that memory to make better decisions over time. AgentDNA is our attempt to create a reliable long-term memory layer for agentic systems, combining structured operational memory with semantic context so agents can evolve instead of starting from zero every time.

What it does

AgentDNA gives AI agents persistent, contextual, and actionable memory.

It captures important interactions, decisions, outcomes, preferences, and experiences, then retrieves the most relevant memories when the agent encounters a similar situation.

The system can:

Store long-term agent memories Retrieve semantically relevant past experiences Maintain structured user and session context Track decisions and their outcomes Use previous experiences to improve future responses Separate short-term conversational context from durable long-term memory Provide explainable memory provenance so agents can understand why a memory was retrieved

The goal is to transform an agent from:

“What should I do now?”

into:

“What have I learned from similar situations before, and how should that influence what I do now?”

How we built it

AgentDNA is designed as a modular agentic-memory architecture.

Core stack:

LangGraph for stateful agent orchestration LangChain for LLM and retrieval workflows CockroachDB for durable, globally distributed structured memory Vector/semantic retrieval for contextual memory AWS services for scalable infrastructure FastAPI for the backend service layer MCP for standardized tool connectivity React/Streamlit frontend for memory visualization and interaction

The workflow follows:

User → Agent → Memory Retrieval → Reasoning → Action → Outcome → Memory Consolidation → Persistent Memory

Built With

• amazon-web-services
• cockroachdb
• docker
• fastapi
• langchain
• langraph
• python
• statesgraph