---

name: online_stream_memory description: Build bounded online continual-learning streams, replay memories, and forgetting ledgers for MIR-style experiments.

Online Stream and Replay Memory

Use this skill when implementing or checking an online continual-learning experiment in the style of Maximally Interfered Retrieval. It is appropriate for bounded recovery runs, deterministic proxy experiments, and tests of replay-memory behavior. Do not use it to score MIR interference; that belongs to the virtual-update interference skill.

Inputs

• Stream examples with features, label, task_id, and stable example_id fields.
• Memory capacity and candidate subset size.
• Optional seed for deterministic candidate sampling and replacement.
• Task-checkpoint accuracy values when computing forgetting.

Outputs

• A bounded ReplayMemory state.
• Candidate replay examples sampled from current memory.
• A forgetting summary with per-task best, final, and forgetting values.

Workflow

1. Initialize ReplayMemory with a fixed capacity and seed.
2. Before each online update, call candidates(candidate_count) to expose a bounded subset of stored examples.
3. After the model update, call add_many(incoming_examples) so memory changes after replay selection.
4. Record task accuracies after task boundaries or final evaluation.
5. Call compute_forgetting(history) to report average forgetting across tasks.

Validation

Run python -m pytest tests or validate the skill tree with validate_skill_tree.py --run-tests. The tests use only the Python standard library and deterministic data.

Limitations

The helper implements deterministic bounded replacement suitable for recovery and smoke tests. It preserves the online memory contract but is not a full high-throughput dataloader or framework-specific replay buffer.