Protecting old knowledge is only half the job

Proposed research direction · November 2026

The fence is plumbing

A fence is any rule that limits which parts of the network an update may change: Subspace Tuning’s projection, a gradient projection, an importance penalty. It is linear algebra, and linear algebra can only decide where an update goes. It cannot decide what deserves protecting, when new learning should be locked in, or what experience should arrive. Those three answers have to come from outside the fence: from memory systems, learning rules and the data stream. The claim to test is that the answers matter more than the fence.

Proposal only. The methods named below are published work by others, cited as examples of each source; none has been run on my agents.

1. Data stream → what flows in

   Own play: on-policy reinforcement learning forgot less than supervised fine-tuning at matched accuracy (RL’s Razor, ICLR 2026). Replay: interleaving a small buffer of earlier experience into the stream (Rolnick et al., 2019), the same interleaving the brain is thought to do during sleep. Order: a gated level chain decides what arrives when, and blocked and interleaved orders forget differently.

2. Memory systems → what to protect

   Gradient Projection Memory stores the directions that earlier tasks’ activations actually used and keeps later updates orthogonal to them: protection measured from memory, not guessed from the weights as Subspace Tuning does (Saha et al., ICLR 2021). Elastic Weight Consolidation remembers how much each weight mattered to earlier tasks and resists changing those (Kirkpatrick et al., 2017). An episodic buffer can check, at update time, that stored old examples do not get worse (A-GEM, ICLR 2019).

3. Learning rules → when to lock in

   Complementary learning systems: the brain learns fast in one system and consolidates slowly into another through interleaved replay, a rule for when new learning moves behind the fence (McClelland, McNaughton & O’Reilly, 1995). Synaptic Intelligence tracks each weight’s importance online during training, so protection updates continuously instead of on a schedule (Zenke et al., 2017). Nested Learning treats a network as memory modules that update at different frequencies (Behrouz et al., NeurIPS 2025). Subspace Tuning’s own refresh re-measures its fence on a fixed timer, which the paper lists as a limitation.