Comparison | Hindsight
Sibyl vs Hindsight: good milestone recall, high context cost
Hindsight retrieves some categories, but sends much more context to the model and remains costly on ingestion.
Full test note
Hindsight is the subtle page. It does not collapse like a bad retrieval toy. It finds meaningful memories. The problem is that it often sends so much context that Sonnet becomes the expensive part of the memory system.
Read the full test pageVerdict
Hindsight passes 152/350 in retrieval and answers, with 11,892 average context tokens and 98 minutes of ingestion.
Hindsight runs as a local daemon, LLM provider `none`, consolidation disabled, recall with max_tokens=4000 and include_chunks=false.
Comparison table
| Metric | Sibyl | Hindsight |
|---|---|---|
| Retrieval | 350/350 | 152/350 |
| Sonnet answers | 344/350 | 152/350 |
| Average context | 228 tokens | 11,892 tokens |
| Ingestion | 48 s | 5,870 s |
| Estimated answer cost | $0.64 | $18.68 |
Methodological caveat
Hindsight is configured without an LLM provider or consolidation. The comparison measures the cost and quality of the tested local mode, not an optimized cloud configuration.
Why not the best possible configuration?
- Enabling an LLM provider would move ingestion into the same cost problem as Mem0.
- Even without a provider, read cost is already high: average context exceeds 11k tokens.
- The page therefore keeps the result as a measure of efficiency for the tested local mode, not as a definitive judgment of the product.
Rerun the test
python scripts/run_hindsight_365d_500c_category_baseline.py