Oracle AI Agent Memory

Persistent memory for enterprise AI agents on Oracle AI Database.

Oracle AI Agent Memory helps agents retain context across interactions, compact active conversations, and retrieve durable user, agent, and thread memories when they matter. It is designed for Python applications that need persistent memory backed by Oracle AI Database.

Get Started

Install the package:

pip install oracleagentmemory

For OCI Generative AI reranking, install the optional OCI SDK dependency:

pip install "oracleagentmemory[rerank-oci]"

Do you need a local Oracle AI Database for development? See how to run Oracle AI Database locally: https://docs.oracle.com/en/database/oracle/agent-memory/26.4/agmea/run-locally.html

Python quickstart

Create a memory client, store a short conversation, and search it back:

from oracleagentmemory.apis.searchscope import SearchScope
from oracleagentmemory.core import OracleAgentMemory
from oracleagentmemory.core.embedders.embedder import Embedder
from oracleagentmemory.core.llms.llm import Llm

embedder = Embedder(model="YOUR_EMBEDDING_MODEL")
llm = Llm(model="YOUR_LLM")
db_pool = ...  # an oracledb connection or connection pool

memory = OracleAgentMemory(connection=db_pool, embedder=embedder, llm=llm)

messages = [
    {
        "role": "user",
        "content": "Orange juice has become my favorite breakfast drink lately. What can I pair it with?",
    },
    {
        "role": "assistant",
        "content": "Try eggs and toast, avocado toast, or a breakfast sandwich.",
    },
]

thread = memory.create_thread(user_id="user_123")
# add_messages will add messages to the DB and extract memories automatically
thread.add_messages(messages)
# add_memory adds memory to the DB
thread.add_memory("The user likes orange juice with breakfast.")

results = memory.search(query="orange juice", scope=SearchScope(user_id="user_123"))
for result in results:
    print(f"- [{result.record.record_type}] {result.content}")

OracleAgentMemory uses your Oracle Database connection or pool as the persistence layer, an embedder for vector retrieval when needed, and an optional LLM for memory extraction, summaries, and context cards. When you know your embedding model's output dimension, you can pass embedding_dimension to Embedder so DB schema setup does not need to infer it with a provider call.

If you want Oracle-managed hybrid search over stored search text instead of vector-only retrieval, opt in when constructing the client:

from oracleagentmemory.core import (
    SearchIndexSyncMode,
    OracleAgentMemory,
    SchemaPolicy,
    SearchStrategy,
)
from oracleagentmemory.core.embedders import OracleDBEmbedder

db_embedder = OracleDBEmbedder(
    connection=db_pool,
    model="YOUR_DB_EMBEDDING_MODEL",
    embedding_dimension=384,
)

hybrid_memory = OracleAgentMemory(
    connection=db_pool,
    embedder=db_embedder,
    llm=llm,
    schema_policy=SchemaPolicy.CREATE_IF_NECESSARY,
    search_strategy=SearchStrategy.HYBRID,
    search_index_sync=SearchIndexSyncMode.ON_COMMIT,
)

SchemaPolicy.CREATE_IF_NECESSARY lets the SDK add the managed search objects needed by the selected strategy. SearchStrategy.HYBRID creates or validates Oracle's managed hybrid vector index and requires the main embedder to be an OracleDBEmbedder; the index uses that embedder's in-database model. search_index_sync controls refresh timing for keyword and hybrid indexes. SearchIndexSyncMode.AUTO is supported only for hybrid search; keyword search supports ON_COMMIT and MANUAL.

Warning: The first creation of a hybrid index over existing data can be long-running because Oracle scans stored search text and builds the managed index state during schema setup. Run that upgrade as a planned migration for large schemas.

SearchStrategy.KEYWORD uses text search over stored search text and does not need an embedder. Keyword schemas created without local vector storage cannot be reopened with SearchStrategy.VECTOR unless you recreate the schema or backfill local embeddings first. They can still be upgraded to hybrid search with an OracleDBEmbedder, because Oracle's managed hybrid index builds from stored search text.

LongMemEval Results

Benchmark	Score	Correct
LongMemEval	94.4	472 / 500
Knowledge Update	94.0	73 / 78
Multi-session Reasoning	88.0	117 / 133
Temporal Reasoning	98.0	130 / 133
Single-session User	97.0	68 / 70
Single-session Assistant	100.0	56 / 56
Preference	93.0	28 / 30

Configuration note:

LongMemEval full 500-question evaluation. Answering with GPT-5.5 using xhigh reasoning effort. Mean answer LLM total tokens: 29,042; mean thinking tokens: 1,178.

BEAM Results

	Score			Correct
	100K	1M	10M	100K	1M	10M
abstention	71%	65%	62%	28.50 / 40	45.50 / 70	12.50 / 20
contradiction_resolution	76%	69%	54%	30.25 / 40	48.00 / 70	10.88 / 20
event_ordering	22%	25%	24%	8.93 / 40	17.42 / 70	4.77 / 20
information_extraction	88%	83%	62%	35.17 / 40	58.19 / 70	12.50 / 20
instruction_following	81%	79%	70%	32.50 / 40	55.00 / 70	14.00 / 20
knowledge_update	60%	66%	70%	24.00 / 40	46.50 / 70	14.00 / 20
multi_session_reasoning	67%	63%	14%	26.81 / 40	43.93 / 70	2.90 / 20
preference_following	88%	82%	56%	35.33 / 40	57.46 / 70	11.25 / 20
summarization	59%	66%	50%	23.59 / 40	46.42 / 70	10.02 / 20
temporal_reasoning	72%	59%	39%	29.00 / 40	41.08 / 70	7.75 / 20
Total	68.5%	65.6%	50.3%	274.08 / 400	459.50 / 700	100.56 / 200

Metric	100K	1M	10M
Mean Input Tokens	64,936	86,417	93,161
Mean Output Tokens	2,117	2,424	2,216
Mean Thinking Tokens	1,850	2,056	1,979

Answering with GPT-5.5 using xhigh reasoning effort.

Why Oracle AI Agent Memory?

Oracle AI Agent Memory is built for teams that want persistent agent memory on Oracle AI Database. It helps applications move beyond single-turn prompts by storing conversation history, durable memories, and prompt-ready short-term context in Oracle AI Database.

It gives agents a consistent way to:

• remember useful information across sessions;
• keep active threads compact and context-rich;
• retrieve memories by user, agent, and thread scope;
• use LLM-backed extraction when automatic memory formation is useful;
• write explicit memories when deterministic application control is preferred;
• integrate memory into existing Python agent applications.

Core Concepts

Threads

A thread represents an ongoing conversation or task. Threads can store user and assistant messages, return recent messages, and produce short-term context for an agent prompt.

thread = memory.create_thread(user_id="user_123")
thread.add_messages(
    [
        {"role": "user", "content": "Remember that I like morning meetings."},
        {"role": "assistant", "content": "I will keep that in mind."},
    ]
)

Durable Memories

A durable memory is information intended to survive beyond the active thread. Applications can add durable memories directly.

thread.add_memory("The user prefers morning meetings.")

When an LLM is configured and memory extraction is enabled, Oracle AI Agent Memory can also extract durable memories from thread messages.

Search Scope

Search is scoped explicitly so applications can control which memories are eligible for retrieval.

thread.add_memory(
    "The user prefers morning meetings.",
    metadata={"source": "support"},
)
results = memory.search(
    query="When does the user prefer meetings?",
    scope=SearchScope(user_id="user_123"),
)

Use record_types and metadata_filter when retrieval should only consider specific kinds of records or records whose stored metadata contains a requested partial mapping. Multiple filter keys are combined with AND semantics, nested dictionaries match nested metadata fields, and list values must match exactly:

results = memory.search(
    query="When does the user prefer meetings?",
    scope=SearchScope(user_id="user_123"),
    record_types=["memory"],
    metadata_filter={"source": "support"},
)

For array membership, put a field-level operator dictionary where an exact value would normally appear. $array_contains with a scalar checks one array value, $array_contains with a list requires all listed values, and $array_contains_any requires at least one listed value. $not negates another field-level operator dictionary; a standalone negated membership filter also matches records where that field is missing or is not an array:

results = memory.search(
    query="When does the user prefer meetings?",
    scope=SearchScope(user_id="user_123"),
    record_types=["memory"],
    metadata_filter={
        "source": "support",
        "tags": {"$array_contains": "support"},
    },
)

Context Cards

For longer conversations, get_context_card() creates prompt-ready short-term context with a summary, relevant durable records, retrieval topics, and recent messages.

context = thread.get_context_card()

Use context cards when a model needs continuity without receiving the full conversation transcript.

Security and Deployment Notes

Oracle AI Agent Memory uses the database connection, model providers, credentials, and network configuration supplied by the integrating application. For production deployments:

• use encrypted database connections and secure model-provider endpoints;
• keep secrets out of source code and checked-in configuration;
• apply end-user authentication and authorization before memory operations;
• pass correct memory scope values for every request;
• bound message sizes, retrieval sizes, and provider usage for your workload;
• review LLM-backed extraction and summarization carefully before using them with sensitive data.

Troubleshooting

If pip install oracleagentmemory or pip install oracleagentmemory==26.4.0 reports No matching distribution found, first check the Python interpreter used by pip:

python --version
python -m pip --version
python -m pip install oracleagentmemory

Oracle AI Agent Memory supports Python 3.10 through 3.14. Python 3.9 can produce a generic resolver error instead of a clear "requires Python >=3.10" message. Create a Python 3.10, 3.11, 3.12, 3.13, or 3.14 environment and install again.

Oracle AI Agent Memory requires vector memory to be configured in Oracle Database before using vector search or vector-index backed schemas. If the database raises ORA-51962: The vector memory area is out of space for the current container, ask a DBA or privileged administrator to size vector memory for both the root container and the target pluggable database. The exact values depend on your database and workload; this example configures 512M at the root and 256M for the PDB:

ALTER SESSION SET CONTAINER = CDB$ROOT;
ALTER SYSTEM SET vector_memory_size = 512M SCOPE=SPFILE SID='*';
SHUTDOWN IMMEDIATE;
STARTUP;
ALTER PLUGGABLE DATABASE <PDB_NAME> OPEN;
ALTER SESSION SET CONTAINER = <PDB_NAME>;
ALTER SYSTEM SET vector_memory_size = 256M SCOPE=BOTH;
SELECT value FROM v$parameter WHERE name = 'vector_memory_size';

See the Oracle Database error help for ORA-51962.

If you need a read-only runtime database user, create or upgrade the managed schema with an owner account, grant only SELECT on the managed tables, and start the runtime client with SchemaPolicy.REQUIRE_EXISTING. The runtime user also needs the normal database privilege required to connect, such as CREATE SESSION, which usually comes from a DBA or privileged administrator.

-- Run as a DBA or privileged administrator.
GRANT CREATE SESSION TO memory_r;

The example below assumes embedder is already configured for your application.

import os

import oracledb

from oracleagentmemory.core import OracleAgentMemory, SchemaPolicy

RUNTIME_DB_USER = os.environ.get("ORACLE_MEMORY_RUNTIME_DB_USER", "memory_r")
TABLE_NAME_PREFIX = "APP_MEMORY_"

runtime_pool = oracledb.SessionPool(
    user=RUNTIME_DB_USER,
    password=os.environ["ORACLE_MEMORY_RUNTIME_DB_PASSWORD"],
    dsn=os.environ.get("ORACLE_MEMORY_DB_CONNECT_STRING", "localhost:1521/FREEPDB1"),
)

memory = OracleAgentMemory(
    connection=runtime_pool,
    embedder=embedder,
    memory_extraction_config=MemoryExtractionConfig(extract_memories=False),
    schema_policy=SchemaPolicy.REQUIRE_EXISTING,
    memory_store_id=memory_store_id,
)

If another agent framework reports dependency conflicts around LiteLLM, OpenAI, or python-dotenv, upgrade to the latest Oracle AI Agent Memory release available to you. Oracle AI Agent Memory 26.5.0 uses litellm>=1.84.0,<2. Avoid installing other packages that pin incompatible LiteLLM ranges in the same environment. For example, if CrewAI's LiteLLM extra selects an incompatible LiteLLM version, install crewai instead of crewai[litellm]; otherwise, isolate the two dependency sets in separate environments.

License

Oracle AI Agent Memory is dual-licensed under:

• Apache License 2.0
• Universal Permissive License (UPL) 1.0

You may choose either license.