Uploaded October 2026 | Updated November 2026, 2 days ago
Read more here: blogs.oracle.com/developers/image-aware-ai-agent-memory-from-visual-inputs-to-searchable-context
AI agent memory for images: Daniela Pavlenco stores a performance chart with a short description in Oracle AI Agent Memory, finds it later by searching in ordinary words from a fresh Python process, and loads the original image bytes back from Oracle AI Database 26ai, then proves they are the same bytes.
0:00 Where did that chart go?
0:17 One chart through Oracle AI Agent Memory
0:38 1. Store the image with useful text
0:58 2. Close the app, start a fresh process
1:19 3. Find it: search reads text, not pixels
1:40 4. Load the search result
2:03 5. Verify: the same bytes, not a look-alike
2:24 6. Scope and limits
2:48 Recap: save, search, load
In this Oracle Developers Explains episode, Daniela Pavlenco, Developer Advocate at Oracle, follows one performance chart through Oracle AI Agent Memory. Your AI agent saw a chart earlier in a conversation, the conversation has moved on, and now you ask for the chart again. Memory here means information saved outside the current conversation, so the application can retrieve it later, even after the application restarts.
Every step is a real local run of the oracleagentmemory Python package (version 26.8.0) against Oracle AI Database 26ai Free. First, a store_images.py process calls add_image_async to save three charts, each with a short text description, a user identifier, an agent identifier and the image format, and Oracle AI Agent Memory returns an image ID. That Python process then exits. A new Python process searches without the file name or the saved image ID, asking in ordinary words for "response time recovering after a slowdown". The search compares the question with the text stored with each image, not with the pixels, so a useful description matters: details that never reach the text may not be found. In this run the search is a keyword search over the stored description.
The matching record returns an image ID, and a third Python process passes that returned ID to list_images_async with include_bytes=True to load the original 22,548 bytes. The SHA-256 fingerprint of the retrieved bytes matches the SHA-256 fingerprint of the original file, and a SQL query computes the same SHA-256 inside Oracle AI Database 26ai over the stored BLOB. A thumbnail or a non-empty record alone would not prove that. The user and agent identifiers organise this example's memory scope, but they are not an access-control boundary: your application still needs identification and access controls. And this is retrieval when your code asks for it, not an agent that automatically remembers the right image in every conversation.
The Python package shown at 0:17 installs with pip install oracleagentmemory (version 26.8.0 on PyPI): pypi.org/project/oracleagentmemory
Why the search at 1:19 reads the description rather than the pixels is covered in What's New in Oracle AI Agent Memory, the 26.8 release overview: medium.com/oracledevs/whats-new-in-oracle-ai-agent-memory-graph-aware-retrieval-image-memory-and-enterprise-controls-2f4ec84d3aca
To try it yourself, the link at 3:02 is Daniela Pavlenco's article Image-Aware AI Agent Memory: From Visual Inputs to Searchable Context on the Oracle Developers blog, with the companion notebook: blogs.oracle.com/developers/image-aware-ai-agent-memory-from-visual-inputs-to-searchable-context
The first episode, AI Agent Memory Explained, covers what an AI agent should keep between sessions: youtube.com/watch?v=YbjDFwMAiuo
Read the blog and run the companion notebook, then tell us in the comments: what is one image your AI agent should be able to find again? Subscribe to Oracle Developers for the next Developers Explains episode.
#AIAgents #AgentMemory #OracleAI
Read more here: blogs.oracle.com/developers/image-aware-ai-agent-memory-from-visual-inputs-to-searchable-context
AI agent memory for images: Daniela Pavlenco stores a performance chart with a short description in Oracle AI Agent Memory, finds it later by searching in ordinary words from a fresh Python process, and loads the original image bytes back from Oracle AI Database 26ai, then proves they are the same bytes.
0:00 Where did that chart go?
0:17 One chart through Oracle AI Agent Memory
0:38 1. Store the image with useful text
0:58 2. Close the app, start a fresh process
1:19 3. Find it: search reads text, not pixels
1:40 4. Load the search result
2:03 5. Verify: the same bytes, not a look-alike
2:24 6. Scope and limits
2:48 Recap: save, search, load
In this Oracle Developers Explains episode, Daniela Pavlenco, Developer Advocate at Oracle, follows one performance chart through Oracle AI Agent Memory. Your AI agent saw a chart earlier in a conversation, the conversation has moved on, and now you ask for the chart again. Memory here means information saved outside the current conversation, so the application can retrieve it later, even after the application restarts.
Every step is a real local run of the oracleagentmemory Python package (version 26.8.0) against Oracle AI Database 26ai Free. First, a store_images.py process calls add_image_async to save three charts, each with a short text description, a user identifier, an agent identifier and the image format, and Oracle AI Agent Memory returns an image ID. That Python process then exits. A new Python process searches without the file name or the saved image ID, asking in ordinary words for "response time recovering after a slowdown". The search compares the question with the text stored with each image, not with the pixels, so a useful description matters: details that never reach the text may not be found. In this run the search is a keyword search over the stored description.
The matching record returns an image ID, and a third Python process passes that returned ID to list_images_async with include_bytes=True to load the original 22,548 bytes. The SHA-256 fingerprint of the retrieved bytes matches the SHA-256 fingerprint of the original file, and a SQL query computes the same SHA-256 inside Oracle AI Database 26ai over the stored BLOB. A thumbnail or a non-empty record alone would not prove that. The user and agent identifiers organise this example's memory scope, but they are not an access-control boundary: your application still needs identification and access controls. And this is retrieval when your code asks for it, not an agent that automatically remembers the right image in every conversation.
The Python package shown at 0:17 installs with pip install oracleagentmemory (version 26.8.0 on PyPI): pypi.org/project/oracleagentmemory
Why the search at 1:19 reads the description rather than the pixels is covered in What's New in Oracle AI Agent Memory, the 26.8 release overview: medium.com/oracledevs/whats-new-in-oracle-ai-agent-memory-graph-aware-retrieval-image-memory-and-enterprise-controls-2f4ec84d3aca
To try it yourself, the link at 3:02 is Daniela Pavlenco's article Image-Aware AI Agent Memory: From Visual Inputs to Searchable Context on the Oracle Developers blog, with the companion notebook: blogs.oracle.com/developers/image-aware-ai-agent-memory-from-visual-inputs-to-searchable-context
The first episode, AI Agent Memory Explained, covers what an AI agent should keep between sessions: youtube.com/watch?v=YbjDFwMAiuo
Read the blog and run the companion notebook, then tell us in the comments: what is one image your AI agent should be able to find again? Subscribe to Oracle Developers for the next Developers Explains episode.
#AIAgents #AgentMemory #OracleAI










