Skip to content

vs. FadeMem

With deep recall + re-ranking (Run H, conv 0):

FadeMem is a research system that introduces temporal decay to memory retrieval. It applies a forgetting curve to penalize old memories during retrieval. Key characteristics:

  • Implements basic Ebbinghaus-style decay
  • Decays memories based on age and importance
  • No retrieval-based strengthening (no spaced repetition)
  • No associative linking
  • No tiered storage or consolidation

FadeMem applies decay but doesn’t strengthen memories when they’re retrieved. In Cognitive Memory, every retrieval boosts stability, creating a spaced repetition dynamic. Important memories that are frequently accessed become increasingly resistant to decay.

FadeMem’s decay is monotonic — once a memory starts fading, it can only get weaker. Cognitive Memory’s decay is non-monotonic — retrieval can reverse the trend.

FadeMem retrieves memories independently. Cognitive Memory creates and maintains an association graph through synaptic tagging and co-retrieval strengthening. This is the primary driver of the multi-hop advantage.

FadeMem has no concept of protected memories. All memories decay at the same rates (modulated by type and importance). Cognitive Memory’s core memories have a 0.60 retention floor, ensuring identity-critical facts are always available.

FadeMem keeps all memories in a single store. Cognitive Memory migrates faded memories to cold storage, reducing search noise. Cold memories can still be activated through associations, preserving multi-hop paths.

FadeMem doesn’t consolidate fading memories. Over time, the memory store accumulates many faded-but-present memories that add noise to search results. Cognitive Memory compresses groups of fading memories into summaries, maintaining signal quality.

When consolidation compresses memories, Cognitive Memory preserves the originals in cold storage accessible via deep recall. FadeMem has no equivalent mechanism.

FadeMem demonstrates an important insight: decay alone isn’t enough. Applying forgetting curves to memory retrieval is a good idea, but without complementary mechanisms (strengthening, associations, protection, consolidation), the improvement over naive RAG is modest.

FadeMem’s published LoCoMo results (27.2% overall) are only marginally better than naive RAG (~20%). The decay model helps with noise reduction, but without associations for multi-hop reasoning and strengthening for importance tracking, the gains plateau quickly.

Cognitive Memory’s approach treats decay as one component of a complete memory lifecycle: encode -> decay -> strengthen -> associate -> consolidate -> recall.

We ran a direct replication of FadeMem’s settings (Run B) to verify our baseline:

ConfigOverallMulti-hop
FadeMem published27.2%29.4%
Our Run B (FadeMem settings)27.7%24.6%

Results are comparable (within noise), confirming our evaluation methodology is consistent.

  • FadeMem results are from their published paper
  • We replicated their settings in our framework (Run B) for verification
  • Both systems use the same LoCoMo dataset and evaluation protocol
  • The comparison is between FadeMem’s published numbers and our Run F (best full-10-conv result)