AI Memory Fails Before It Even Tries — Here’s Why It Matters
If you use AI assistants, this quiet failure could slow down your work without you knowing.
Agentic memory under a fixed budget involves two stages: retention and retrieval. Existing retrieval-centered paradigms implicitly assume necessary evidence survives eviction, but we challenge this by isolating a pre-retrieval failure mode: structurally indirect prerequisite eviction, in which upstream blocks weakly aligned with the query are discarded under budget pressure.
We provide an operational definition of this failure, a reproducible deterministic benchmark, and per-seed trace diagnostics. Finally, we evaluate Dependency-aware Semantic Garbage Collection (DSGC), a one-hop graph-aware rule.
In our main suite, DSGC improves full-chain retention from 0.03 to 0.90 under a lexical encoder and from 0.23 to 1.00 under a sentence encoder. Robustness checks then identify the budget and scaling regimes where the one-hop rule holds or degrades. Our released pipeline and failure postmortem support mechanistic analysis of retention before retrieval as a distinct failure boundary.
- AI assistants can forget important details before they even look them up, making answers less reliable.
- A new fix ‘DSGC’ improved memory retention from 3% to 90% in tests by changing how AI decides what to keep.
- The problem affects everyday AI tools, but the fix only works for simple chains of memory, not complex ones.
Why It Matters
Your AI assistant might be forgetting key details before it even starts searching, making its answers less trustworthy when you need them most.