integrations/pydantic-ai/v0.1.0

moorcheh-ai/memantointegrations/pydantic-ai/v0.1.0Oct 2, 2026by het0814

Upgrading?

AI Summary

Initial release of the Pydantic AI integration providing persistent memory and automatic recall for Pydantic agents.

Key Highlights

  • Automatic memory recall injection into agent instructions.
  • Three explicit tools: memanto_remember, memanto_recall, memanto_answer.
  • Agent lifecycle helper to create and activate agents.
  • Retry logic for invalid tool arguments.

New Features

  • Automatic recall functionality (memory_instructions).
  • Memory tools creation (create_memanto_tools).
  • Agent lifecycle helper (MemantaSetup).
  • Packaging support for Python 3.10+ and pydantic-ai-slim.

Full Release Notes

# Release Notes for memanto-pydantic-ai v0.1.0

Initial release of `memanto-pydantic-ai`, giving [Pydantic AI](https://ai.pydantic.dev) agents persistent, cross-session memory backed by Memanto. Relevant memories are injected into each run automatically, and three tools let the model save, search, and query memory explicitly.

## New Features
- **Automatic recall** (`integrations/pydantic-ai/memanto_pydantic_ai/tools.py`)
  - `memory_instructions(client, agent_id)` plugs into `Agent(instructions=...)`. Before each run it recalls the memories most relevant to the user's prompt and adds them to the agent's instructions.
  - Runs once per agent run (cached across tool-call steps). Only active memories are injected, never expired ones.
  - If the recall fails, it logs a warning and the run continues without memories, so a memory outage never takes the agent down.
- **Memory tools** (`integrations/pydantic-ai/memanto_pydantic_ai/tools.py`)
  - `create_memanto_tools(client, agent_id)` returns `memanto_remember`, `memanto_recall`, and `memanto_answer` for `Agent(tools=...)`.
  - `include_remember` / `include_recall` / `include_answer` select a subset, e.g. for a read-only agent.
  - Invalid tool arguments (unknown memory type, out-of-range confidence or limit, over-long title) are sent back to the model as retries instead of failing the run.
- **Agent lifecycle helper** (`integrations/pydantic-ai/memanto_pydantic_ai/tools.py`)
  - `MemantoSetup()` creates the Memanto agent if it doesn't exist and activates a session. It reads `MOORCHEH_API_KEY` when no key is passed.
  - Raises a clear `ValueError` at startup when a client isn't bound to the right agent, instead of failing on the first tool call.
- **Packaging** (`integrations/pydantic-ai/pyproject.toml`)
  - Requires Python 3.10+, `pydantic-ai-slim>=2.0.0` (no bundled model-provider SDKs) and `memanto>=0.2.21`.

## Tests
- `integrations/pydantic-ai/tests/test_tools.py`: 40 tests that run a real Pydantic AI agent against a mocked client. They cover setup, tool behaviour, retry-on-invalid-input, failing early on a misconfigured client, and automatic recall (caching, outage handling, multimodal prompts, query truncation).

## Full Changelog
Full Changelog: https://github.com/moorcheh-ai/memanto/tree/integrations/pydantic-ai/v0.1.0/integrations/pydantic-ai

Related Repositories