oracle-ai-developer-hub
Agent Memory
A collection of notebooks demonstrating how to build memory-augmented AI agents on top of Oracle AI Database as the unified memory core for AI agents.
What is Agent Memory?
Agent memory is what separates a stateless LLM call from an agent that learns, adapts, and stays coherent across turns, sessions, and runs. It is the substrate that lets an agent:
- Recall prior conversations and reuse what was already said or decided
- Remember durable facts about users, entities, tools, and the world
- Reason over its own past behavior — what it tried, what worked, what failed
- Resume workflows after a crash, restart, or hand-off — without losing state
Practical agent memory is rarely a single store. It is a small set of access patterns layered over the same backend:
| Memory type | What it holds | Typical access pattern |
|---|---|---|
| Conversation / thread memory | Turn-by-turn dialogue history for a single run | Append + ordered read |
| Episodic memory | Discrete events the agent participated in | Time-ranged search |
| Semantic memory | Durable facts and learned knowledge | Vector + keyword search |
| Procedural memory | How-to knowledge, routines, tool-use patterns | Lookup by task |
| Working memory | Scratchpad for the current step | Read/write within a run |
| Entity memory | Facts scoped to a user, customer, or object | Scoped queries |
Oracle AI Database — the unified memory core
Oracle AI Database is the unified memory core for AI agents. Rather than stitching together a vector DB, a key-value store, a graph DB, and a relational store — each with its own client, consistency model, and ops surface — Oracle AI Database serves all of these access patterns from a single converged engine:
- Vector search for semantic recall and retrieval-augmented generation
- Relational queries for structured agent state and audit trails
- JSON / document for flexible message and tool-call payloads
- Graph for relationships between entities, sessions, and findings
- Spatial and full-text for richer retrieval over the same memories
The notebooks in this folder use the oracleagentmemory (OAMP) Python package, which is the AI-Agent Memory Package built on top of Oracle AI Database. OAMP wraps the database as a memory backend with a consistent API for users / agents, memories, and threads — the three primitives behind every notebook in this folder.
Notebooks
| # | Name | Description | Framework | Open Notebook | Open in Colab |
|---|---|---|---|---|---|
| 01 | Deep Research Agent | Build a deep research agent for human genome exploration that uses Tavily for live web search and stores running conversation + durable findings in Oracle AI Database. Demonstrates the OpenAI Agents SDK Session protocol implemented against an Oracle-backed memory store. |
OpenAI Agents SDK · Tavily · OAMP |  |
 |
| 02 | Supply Chain Assistant | A supply chain assistant that tracks and updates shipment cargo through in-process tools and an MCP server. Uses Oracle AI Agent Memory to persist shipment records, operational notes, and conversation history across restarts. | Claude Agent SDK · MCP · OAMP | |
|
| 03 | Mortgage Approval Workflow | A deterministic mortgage approval workflow modeled as a StateGraph with prebuilt create_agent nodes. Uses Oracle AI Agent Memory so a workflow that fails mid-stage can resume from the last persisted state instead of restarting. |
LangGraph · OAMP | |
|
| 04 | OAMP Benchmarks | Quantifies the practical benefits of Oracle AI Agent Memory over naive flat-history memory along three axes: token consumption per turn, wall-clock latency, and response quality (LLM-as-a-judge). Runs the same 80-turn scripted conversation through three agents. | OAMP · LiteLLM · OpenAI | |
|
| 05 | OAMP Developer Guide | A hands-on, step-by-step guide to the oracleagentmemory package. Builds an agent memory system from scratch — connection, the three primitives (users/agents, memories, threads), manual vs. automatic LLM-powered extraction, vector search, context cards, and scoping. |
OAMP · LiteLLM | |
|
| 06 | Support Assistant Copilot | An end-to-end customer-support copilot that follows one damaged-delivery case from setup through knowledge ingestion, agent tool use, context-card compaction, preference correction, cross-user isolation, TTL/retention, and teardown. Demonstrates background extraction, pluggable embeddings (OpenAI by default or an in-database ONNX model), vector search, metadata inheritance and filtering, and chunked semantic indexing. | OpenAI Agents SDK · OAMP | |
|
| 07 | Short-Term vs Long-Term Agent Memory | Lifecycle walkthrough for active thread state, summaries, typed durable memory, metadata and thread scope, context cards, updates, TTL, and deletion in Oracle AI Agent Memory 26.6. | OAMP 26.6 | |
|
| 08 | Agent Memory Checklist | The agent memory checklist proved live on Oracle AI Database 26ai, with Claude as the agent. For each capability (in-database embedding, vector, keyword and hybrid recall, filtering, reranking, graph, ACID corrections, Flashback history, JSON Relational Duality, files and a DBFS scratchpad, Deep Data Security, auditing) Claude answers the same question without and with it. Setup, data, models and diagrams all live in the notebook; the appbook runs the same checklist in the browser. | Anthropic SDK · Oracle 26ai | |
|
Getting Started
If you are new to Oracle AI Agent Memory, the recommended order is:
- Start with the Developer Guide (05) — learn the API surface and the three core primitives.
- Run the Benchmarks (04) — see the cost, latency, and quality differences vs. naive memory.
- Pick a framework example — OpenAI Agents SDK (01), Claude Agent SDK (02), or LangGraph (03), depending on your stack.
Prerequisites
- An Oracle AI Database instance — either local (Docker) or reachable over the network
- An LLM provider key (
OPENAI_API_KEY,ANTHROPIC_API_KEY, or any LiteLLM-supported provider depending on the notebook) - Python 3.10+ with the
oracleagentmemorypackage installed