A Pydantic AI Harness agent can search tables already in your catalog, and it can keep its notes in one of those tables. Use either piece, or both. The Harness guide is Pixeltable. Pydantic models as column values are Working with Pydantic.
Pixeltable(['hr.handbook']) is a capability. The list is an allowlist of table paths or directory prefixes. PixeltableMemoryStore(table_name='hr.memory') is the store you pass to Memory. The store creates that table on first use.
Install
Python 3.11 to 3.14. The pixeltable extra installs Pixeltable 0.7.21 or later.
pydantic-ai-slim[openai] installs the openai package. The example model and pixeltable.functions.openai.embeddings both call it. Export OPENAI_API_KEY before the insert and before run_sync(). Pixeltable does not load a .env file.
Example
In app.py, declare the handbook on a TableModel and the embedding index in __indexes__. Run pxt init once in the project, then pxt schema update app.py hr, which creates hr.handbook. See Embedding indexes. The Harness guide uses create_dir(), create_table(), and add_embedding_index() for the same table in a notebook.
Importing app.py declares the model but does not attach it to the table. Call bind_all() with the same last argument as pxt schema update first, or insert() raises NOT_BOUND. insert() calls OpenAI to embed text.
The first run_sync() creates hr.memory.
A video example indexes transcripts and frame captions, then searches that stored text: pixeltable_video.py.
query_table and similarity_search return {"table", "rows", "truncated"}. The default caps are 20 rows and 8000 characters (max_rows and max_chars on Pixeltable). Results skip media, array, binary, unstored computed columns, and unstored iterator outputs. Naming one of those columns in a projection or a filter is an error, and the tool asks the model to retry. A versioned path such as 'hr.handbook:3' is refused.
['*'] allowlists the whole catalog, including a memory table. Two Pixeltable capabilities on one agent share id='pixeltable'. The same configuration stated twice is one capability. Two that differ in any setting (tables, max_rows, max_chars, guidance) raise UserError. To keep both, give each its own id and wrap them in PrefixTools from pydantic_ai.capabilities.
Writes to PixeltableMemoryStore are compare-and-set, and each mutation keeps an operation receipt, so a retried tool call replays instead of applying twice. Memory(write_metadata=...) raises UserError on this store. A deleted or overwritten memory stays readable at an earlier version, such as pxt.get_table('hr.memory:3'). Drop the table to remove that history. store.table is the Pixeltable table. Filter kind == 'file' for memory rows, and send writes and deletes through the store. A memory write inserts rows, and insert() logs a line that begins Inserted 1 row with 0 errors followed by the elapsed time. Set PIXELTABLE_VERBOSITY=0 to silence it.
While Harness is on 0.x, the API can change between minor releases. Deprecation warnings and the release notes name the replacement. See the version policy.
Pixeltable loads from an agent spec when you pass custom_capability_types=[Pixeltable]. PixeltableMemoryStore stays in Python. The spec backends for Memory are memory, file, and sqlite. Last modified on November 7, 2026
