Free-Recall Net Learns Several Memory Strategies

bravo_abad · x · 2026-07-20

A Nature Machine Intelligence paper finds that a neural net trained for free recall can learn multiple memory strategies, not just the classic temporal-context style.

The setup: a GRU stores snapshots of its hidden state during study, then retrieves items by cosine similarity under reinforcement learning.
Across many runs, behavioral and representational clustering reveals three strategies: memory-palace / fixed-index routing, TCM-like forward, and TCM-like backward.
Which strategy emerges depends on training conditions:
a higher RL discount factor pushes the model toward fixed position codes;
flushing working memory before recall also favors that route;
training noise strengthens the effect and makes the index code more robust at test time.
In a key-value variant, keys encode position while values carry content, even though that split was never explicitly designed.
The result is task-specific: if retrieval is cued by item features instead of position, models do not learn the memory-palace strategy.

The practical takeaway for RAG and agentic systems is that memory behavior is shaped by the objective, horizon, and noise floor; content-based and position-based indexing fail in different ways, so you should know which regime your system is in before deploying it.

Original post →

Free-Recall Net Learns Several Memory Strategies

More from coding & agent