Memory Architecture Research
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
| Dimension | Holographic | Sibyl Memory |
|---|---|---|
| Storage | SQLite + FTS5 (3 tables) | SQLite + FTS5 (5 tiers) |
| Locality | Fully offline | Requires cloud activation |
| Cost | Free, unlimited | Free cap: 2 MB |
| Feature | Trust scoring, HRR vectors | Self-learning, 5-tier schema |
| Search | FTS5 + trust-weighted | FTS5 + multi-tier + LLM re-ranking |
| Dependencies | numpy (optional) | pip + cloud account |
The Verdict
- Sibyl wins on features — self-learning, multi-tier schema, validated retrieval. But the 2 MB free cap and cloud activation requirement are dealbreakers for an offline-first system.
- Holographic wins on operational simplicity — zero dependencies, unlimited storage, no accounts. Less flashy but operationally bulletproof.
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
sibyl_memory/(local) — Sibyl SDK research and evaluationsibyl_memory_evaluation_sources/(local) — Source material for the comparison- Published Blog Post — Full comparison with architecture details and benchmarks
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.