A deep-dive comparison of two SQLite-backed agentic memory systems for Hermes Agent: Sibyl Memory (the #2-ranked system on LongMemEval) and Holographic (our local-first baseline).

The Question

Agents need memory to work across sessions. But every memory system makes a tradeoff: cloud vs local, free vs paid, simple vs feature-rich. Which one is right for an offline-first, zero-tracking research setup?

Key Findings

DimensionHolographicSibyl Memory
StorageSQLite + FTS5 (3 tables)SQLite + FTS5 (5 tiers)
LocalityFully offlineRequires cloud activation
CostFree, unlimitedFree cap: 2 MB
FeatureTrust scoring, HRR vectorsSelf-learning, 5-tier schema
SearchFTS5 + trust-weightedFTS5 + multi-tier + LLM re-ranking
Dependenciesnumpy (optional)pip + cloud account

The Verdict

Where We Landed

We’re staying on Holographic for the orchestrator, research pipelines, and autonomous morning briefing. The self-learning gap is real, but not worth trading local-first operation.

Full Writeup

Sibyl vs Holographic — Memory at the Grain (published on grainworks.tech)

Files

Status → 📖 Research

The comparison is complete and published. Memory architecture remains an active area of interest — particularly the prospect of a fully local port of Sibyl’s multi-tier schema and self-learning detection.