Oracle AI Agent Memory: SDK, Setup, Limits

Oracle AI Agent Memory is Oracle’s memory SDK for AI agents, installed with pip install oracleagentmemory. It stores conversation threads, summaries and long-term memories inside Oracle AI Database, extracts memories from chats with an LLM, and retrieves them with vector, keyword or hybrid search. It suits teams that already run Oracle and want agent memory under the same governance.

What is Oracle AI Agent Memory?

Oracle AI Agent Memory is a model- and framework-agnostic Python library that turns Oracle AI Database into an agent memory store. Oracle’s docs call it “the governed memory core in the AI agents ecosystem.” Short-term memory lives in threads with summaries and context cards. Long-term memory holds user preferences, learned rules and facts across sessions.

The pitch is one database instead of several. Oracle’s About page says it uses Oracle AI Database “as the converged platform for working memory and long-term memory, rather than piecing together separate vector, graph, JSON, and transactional stores.”

Release history on PyPI: 26.4.0 (April 30, 2026), 26.6.0 (July 7, 2026), 26.8.0 (September 22, 2026). The package license is Apache-2.0 OR UPL-1.0, and it supports Python 3.10 to 3.14.

How Oracle AI Agent Memory works

Oracle’s developer blog describes four memory types: working (state the agent reasons over now), semantic (facts about users and the world), episodic (specific past experiences) and procedural (rules and learned procedures). In the SDK they map to a few objects:

• Threads. memory.create_thread(user_id=...) opens a conversation. thread.add_messages(...) stores messages and extracts memories from them automatically.
• Memories. thread.add_memory(...) stores a fact directly. Records carry user, agent, thread and timestamp fields.
• Context cards. get_context_card() returns a compact block with the thread summary, recent messages and relevant memories, so the agent doesn’t resend the full transcript.
• Search. memory.search(query, scope=SearchScope(user_id=...)) returns both memories and raw messages. Three strategies: VECTOR (Oracle AI Vector Search, database 23.4+), KEYWORD (23.4+, no vector index), and HYBRID (23.6+, managed hybrid vector indexes).
• Linked memory. Optional schemas build an Oracle property graph between records (needs CREATE PROPERTY GRAPH privileges).
• Time-to-live. By default nothing expires. MemoryRetentionConfig or per-record TTL sets retention; an optional daily purge job deletes expired rows.

Models go through LiteLLM-style names. The docs list OCI-hosted models (Gemini, Cohere), OpenAI, Anthropic, Gemini and self-hosted gpt-oss for the LLM, and OCI Cohere, OpenAI and Gemini models for embeddings.

How to set up Oracle AI Agent Memory in Python

The steps below follow Oracle’s Get Started guide for version 26.8.

1. Run an Oracle AI Database (23ai or later; the docs link a “Run Oracle AI Database Locally” page).
2. Give the schema owner CREATE TABLE and CREATE PROCEDURE (plus CREATE JOB if you want the purge job).
3. Install the SDK: pip install "oracleagentmemory==26.8.0".
4. Create an embedder, an LLM and a database pool, then the memory object.
5. Open a thread per user and add messages.
6. Search by user scope, or ask for a context card before each model call.

 1import oracledb
 2
 3from oracleagentmemory.core import SchemaPolicy
 4from oracleagentmemory.core.oracleagentmemory import OracleAgentMemory
 5from oracleagentmemory.apis.searchscope import SearchScope
 6from oracleagentmemory.core.embedders.embedder import Embedder
 7from oracleagentmemory.core.llms.llm import Llm
 8
 9embedder = Embedder(model="openai/text-embedding-3-small")
10llm = Llm(model="openai/gpt-4.1")
11db_pool = oracledb.SessionPool(user="AGENT", password="...", dsn="localhost/FREEPDB1")
12
13memory = OracleAgentMemory(
14    connection=db_pool,
15    embedder=embedder,
16    llm=llm,
17    schema_policy=SchemaPolicy.CREATE_IF_NECESSARY,
18    memory_store_id="T_GET_STARTED",
19)
20
21thread = memory.create_thread(user_id="user_123")
22thread.add_messages([
23    {"role": "user", "content": "Orange juice is my favorite breakfast drink lately."},
24    {"role": "assistant", "content": "Try it with eggs and toast."},
25])
26thread.add_memory("The user likes orange juice with breakfast.")
27
28for result in memory.search(query="orange juice", scope=SearchScope(user_id="user_123")):
29    print(f"- [{result.record.record_type}] {result.content}")

Oracle’s blog lists how-to guides for LangGraph, the Claude Agent SDK, the OpenAI Agents SDK, WayFlow and custom Python harnesses. The docs also cover an MCP server for Oracle AI Database Private Agent Factory.

Benchmarks and token savings

Oracle’s developer blog (May 2026, Wojtek Pluta) reports 93.8% on LongMemEval (469 of 500). In an 80-turn test, input stayed at about 1,300 tokens per request versus over 13,900 for a flat-history baseline, roughly 9.5x fewer by the last turn. A judge preferred the memory version in 48 turns, the baseline in 13, with 19 ties.

These are Oracle’s own runs. Every agent memory vendor reports its own LongMemEval score, with different answer models and judges, so a 1-point gap between vendors says little. Oracle’s post doesn’t report the longer BEAM benchmark, where differences between systems are bigger (see the agent memory leaderboard).

Oracle AI Agent Memory vs Hindsight vs Mem0

Hindsight and Mem0 numbers come from Mem0’s own Mem0 vs Hindsight page and Hindsight’s benchmark site. See Mem0 vs Hindsight for that comparison in full.

Where Hindsight fits for Oracle shops

Hindsight can use Oracle AI Database 23ai as its store, so choosing Oracle as the database doesn’t force Oracle’s SDK. The trade is what sits on top. Hindsight runs four retrieval strategies at once and merges them, and it consolidates facts into observations and mental models that reflect reasons over. It also leads on long histories: 73.9 on BEAM 1M and 64.1 on BEAM 10M, 9.8 and 15.5 points above Mem0 on Mem0’s own comparison page.

The limits matter too. Hindsight returns more context per retrieval than Mem0 by default (Mem0 measured 23.9K to 43.6K tokens depending on benchmark), so set max_tokens lower if cost matters. Its writes call an LLM, and it has no published compliance certifications. Oracle’s SDK inherits the database’s auditing and access control, which some regulated teams need. More on Hindsight’s design in Hindsight AI memory explained.

When to choose Oracle AI Agent Memory

• Pick Oracle AI Agent Memory if your data and agents already live in Oracle AI Database and you want memory inside the same backup, replication, encryption and audit setup, or if you build on Oracle’s Private Agent Factory.
• Pick Hindsight for long conversation histories, reasoning over memory, coding-agent plugins, or a fully open server on PostgreSQL. It still runs on Oracle AI Database if that’s your standard.
• Pick Mem0 for a lean per-user memory API with low tokens per retrieval and published SOC 2 and HIPAA.

For the wider field, see open-source memory systems compared and vector databases for LLM memory.