Not Memory, But Forgetting

Claude Code reads CLAUDE.md at session start and loads the entire file directly into context as-is. There's no vector DB or semantic search, and the auto memory in v2.1.59+ leaves notes as markdown files. However, MEMORY.md has a 200-line cap, so anything beyond that is silently truncated.

Mem0 extracts facts from conversations, stores them as embeddings, and retrieves them via semantic search, but it only achieved 49.0% recall on LongMemEval. Token usage is around 1,764 tokens per conversation.

Zep does better at 63.8% on LongMemEval, but it uses about 600,000 tokens per conversation, incurring nearly 340x the memory cost of Mem0. The author adds that the Zep team pushed back, claiming 75.1% under proper configuration.

Letta/MemGPT treats the context window like RAM and external storage like disk. It performs best in benchmarks at about 83.2%, but because it continuously spends reasoning tokens on memory operations themselves, judging "what to remember" comes at a cost.

The core conclusion is simple. Today's agent memory is either

• too simple, so retrieval fails,
• too expensive, so it overconsumes tokens, or
• too smart, so it wastes compute on managing memory.

Ultimately, the problem isn't memory but forgetting — the bottleneck is that we still lack a mechanism, like humans have, to retain only what's important and forget the rest.