Drift: knowledge slowly deviating from the truth

Semantic drift is the headline failure of evolving memory, and both sources describe it. It is caused by lossy compression, especially repeated summarization. Each time information is re-encoded (raw text to summary to higher-level reflection), nuance is stripped away, and over many rounds the memory gradually distorts.

The SSGM paper gives a memorable concrete example: a mild user preference, "I like mild spicy food," gets rewritten over time to "likes spicy food," then "loves very spicy food," until the agent commits a real violation by recommending ghost pepper wings. The drift was slow and each step looked reasonable, which is exactly why it is dangerous.

The "Practical Guide" calls this summarization drift and adds a practical defense: keep raw memories linked to their summarized versions, so you can always reconcile against what actually happened. The SSGM paper formalizes the same defense as anchoring against an immutable ledger, which Chapter 5 covers.

The paper notes two cousins of semantic drift:

• Procedural drift: the agent reinforces a suboptimal or outdated workflow. For example, it "learns" a convoluted workaround for a simple API call and rigidifies it, blocking future improvement.
• Goal drift: the agent's behavior slowly shifts away from its original instructions because of accumulated biases in its memory.

The "Practical Guide" describes the closely related attention dilution: even when you can fit everything in context (as with very large windows), large prompts "lose" information in the middle, so the agent has the memories but cannot focus on the right parts. This is the same "lost in the middle" effect from Chapter 2.