What

Adds opt-in file-backed persistence to the memory plugin, so extracted memories can survive across requests. Closes #111.

Today memory_plugin.run() builds a fresh in-RAM Memory() on every request, so nothing is retained between calls. This adds an optional persistent store, gated behind a single environment variable.

How

• Memory(persist_path=...) loads any previously saved items on init and writes after every add().
• run() reads the path from the new OPTILLM_MEMORY_FILE env var (matching the existing OPTILLM_* env-var convention). When it is unset, persist_path is None and behaviour is byte-for-byte unchanged — the default path stays fully in-RAM and touches no disk.
• Robustness:
  • _load_from_file: a missing file is a normal first run; a corrupt / non-list JSON payload is logged and degrades to an empty store instead of raising, so persistence can never fail a request at startup. Loaded items are filtered to strings and truncated to max_size (most recent kept).
  • _save_to_file: atomic write via a temp file in the same directory + os.replace, so a crash mid-write can't corrupt an existing store. Any OSError is logged and swallowed.
• Growth stays bounded by the existing max_size (100) eviction.

Usage

export OPTILLM_MEMORY_FILE=~/.optillm/memory.json
# memory-gpt-4o-mini requests now load & persist across runs

Design note

The linked issue suggested a JSON save_to_file/load_from_file pair, and you greenlit "add a plugin that supports this." I implemented it as an opt-in extension of the existing memory plugin rather than a separate plugin, to avoid duplicating the extraction/retrieval logic — the default behaviour is untouched. Happy to split it into a standalone plugin instead if you'd prefer that shape.

Tests

tests/test_plugins.py::test_memory_plugin_persistence (runs in the existing no-server unit-test job) covers: round-trip persistence, missing file, corrupt file, non-list payload, max_size truncation on load, and that the default (no persist_path) performs no file I/O.