Memory systems ranked for this Hermes install
Bottom line: for a practical next test, start with Mnemosyne as the baseline, then test Open Second Brain and Ariadne. For service-grade/daemon-backed memory, test Hindsight. Avoid trying to make Honcho or Memory OS the first local target unless the goal is specifically to debug/operate a heavier stack.
Executive ranking
Mnemosyne
Best current fit. Real Hermes MemoryProvider, local SQLite core, mature Hermes docs/tests, MCP optional. Since you have used it before, it is the best baseline and continuity path.
Open Second Brain
Best rich local artifact/vault model. Real Hermes provider, Obsidian/Markdown storage, strong test surface. Main risk is Bun/Node complexity and huge feature surface.
Ariadne
Best new lightweight candidate. SQLite + FTS5 + FAISS + graph, native provider code. Needs Windows/Hermes import fix before serious use.
Hindsight
Best service-grade option. Strong retain/recall/reflect model, REST/MCP/Hermes docs. Requires Docker/Postgres/LLM or managed service.
agent-memory-skill
Best deterministic safety design. SQLite FTS5, authority lanes, provenance, procedural gate. Not a true provider; plugin/hook injection.
Smriti
Clean MCP candidate. Small Python/SQLite/MCP package with temporal facts and procedural skills. Good to test, likely not primary provider.
My opinion after deep research
If the goal is best memory backend for this Hermes install, do not begin with the biggest system. Begin with the system that is provider-shaped, local, and debuggable. That means Mnemosyne first. It has the least conceptual impedance with Hermes, the best maturity signals among Hermes-native local providers, and you already have prior context with it.
Then test Open Second Brain because it is genuinely Hermes-native and stores memory as inspectable Markdown/Obsidian artifacts, which fits your preference for durable local files. After that, test Ariadne because it is small and promising but needs a fix for its hardcoded Hermes import path. Hindsight is worth testing only if you are okay with a background service and LLM-backed extraction.
Provider classification
| Category | Systems | Interpretation |
|---|---|---|
| Real Hermes memory provider | Mnemosyne, Open Second Brain, HyAtlas, Ariadne, SuperLocalMemory/MSLM; Hindsight/Honcho providers appear to live on Hermes side | Can potentially be selected through memory.provider / hermes memory setup. |
| MCP memory system | Smriti, Noshy, Rhizome, Unforgit, Context-MCP, Hindsight, Honcho, SuperLocalMemory | Good tool integration, but may not transparently inject recall unless Hermes/provider glue is added. |
| Plugin/tool/hook memory | agent-memory-skill, remember, hermes-memlock, Memory OS/Icarus | Useful, but not equivalent to a clean provider backend. |
| Library/component | dyadcore, edgeless-memory, remember | Could be wrapped, but not the first choice if a provider already exists. |
Detailed decision matrix
| System | Fit | Why it matters | Main risks | Test priority |
|---|---|---|---|---|
| Mnemosyne | Excellent | Real provider; local SQLite; MCP; BEAM working/episodic memory; sqlite-vec/FTS; strong Hermes docs and adoption signals. | Plugin/package version compatibility; Windows symlink/install docs; sqlite-vec/embedding dimension mismatches. | 1 |
| Open Second Brain | Excellent | Real provider; Markdown/Obsidian vault; deterministic artifacts; many Hermes hooks/tools; large tests. | Bun/Node dependency; large surface area; Windows plugin install path needs smoke test. | 2 |
| Ariadne | Very promising | Native provider code; SQLite + FAISS + FTS5 + graph; local-first, no daemon. | Hardcoded /usr/local/lib/hermes-agent; package naming mismatch; young project. | 3 |
| Hindsight | Strong if service is acceptable | Mature retain/recall/reflect model; REST/MCP; Docker and embedded options; good docs. | Postgres/pgvector or embedded pg0; LLM dependency; async deriver delay; heavier ops. | 4 |
| agent-memory-skill | Great plugin, not provider | Authority lanes, provenance, source policy, audit trail, deterministic SQLite FTS. | Manual install; hardcoded ~/.hermes; not provider-shaped. | 5 |
| Smriti | Good MCP candidate | Small Python/SQLite, temporal facts, procedural skills, offline hash embedder, MCP server. | Young/small; ~ path expansion issue; not native provider. | 6 |
| HyAtlas | Feature-rich but heavy | Real provider entry point, layers L0-L7, Qdrant/Kuzu/dashboard. | Qdrant/Kuzu/hy-memory server complexity; PyPI lag; many moving pieces. | Later |
| Noshy | Good MCP utility | SQLite, MCP/HTTP/dashboard, optional fastembed, memory/memoir concepts. | No real provider; lifecycle/auth/server management. | Later |
| Remember | Simple engine | SQLite/FTS5, optional embeddings, clean tools. | License ambiguity; tool-only integration; very young. | Later |
| Unforgit | Repo memory | Good for developer/repo-scoped memory via MCP. | Node20/better-sqlite3; repo-local model not global profile memory. | Later |
| Rhizome | Knowledge vault | Markdown/DuckDB local wiki compiler, MCP search/temporal/graph. | Not conversational memory; writer-agent workflow; dependency heft. | Later |
| Honcho | Not first local target | Powerful production memory, peer-centric model, Hermes setup docs. | Postgres+pgvector, Redis, API, deriver, LLM config; user already hit local friction. | Only if debugging Honcho specifically |
| Memory OS | Too invasive for first test | Comprehensive Hermes memory operating model. | Whole-stack overlay: Qdrant/Redis/ARQ/Icarus/SOUL/rulebook/cron/wiki. Not a clean provider. | Design inspiration only |
| SuperLocalMemory/MSLM | Risky | Interesting multi-scope local memory; provider code and MCP docs. | AGPL, heavy pinned deps, doc/package inconsistencies, maturity concerns. | Later/avoid embedding |
| dyadcore | Component only | Nice zero-dep SQLite trigram/graph library. | No provider/plugin/MCP; requires custom glue. | Only if building custom provider |
| edgeless-memory | Too experimental | Tiny local SQLite CLI. | No provider/MCP; installer assumes python3; basic schema. | Low |
| context-mcp | Project context only | Simple MCP context store/codegraph. | No tests; JSON concurrency/scaling; license metadata mismatch; not provider. | Low |
| hermes-memlock | Companion only | Useful for pinned standing instructions surviving compaction. | Not factual memory; prompt-injection/pinning risk; concurrency note. | Companion plugin, not provider |
Recommended next test plan
- Baseline: Mnemosyne. Validate install, provider activation, remember/recall, sqlite DB mutation, and cross-session recall.
- Artifact/vault path: Open Second Brain. Validate plugin/provider, vault initialization, Markdown memory artifact creation, and recall context.
- Lightweight new provider: Ariadne. First patch or work around hardcoded Hermes path, then test remember/recall/graph/export.
- Service-grade path: Hindsight. Start Docker/local server, configure Hermes provider, verify retain/recall after deriver processing.
- MCP comparison: Smriti. Add MCP server, test save_fact/recall/save_skill and inspect SQLite.
Concrete smoke commands to start with
Mnemosyne
pip install mnemosyne-hermes
hermes config set memory.provider mnemosyne
hermes memory setup
hermes memory status
# Then in a fresh Hermes session:
# remember: "My smoke-test memory color is teal."
# new session: "What is my smoke-test memory color?"Open Second Brain
hermes plugins install itechmeat/open-second-brain --enable
hermes gateway restart
~/.hermes/plugins/open-second-brain/scripts/o2b install-cli
o2b init --vault C:/Users/Gaem4090/tmp/osb-vault --name "Smoke Brain" --agent-name hermes --timezone UTC
o2b brain init --vault C:/Users/Gaem4090/tmp/osb-vault --primary-agent hermes
o2b doctor --vault C:/Users/Gaem4090/tmp/osb-vault --repo .
hermes memory setup
hermes memory statusAriadne
# First verify package/import naming and patch hardcoded Hermes path if needed.
pip install "ariadne-memory[embeddings]"
python -c "import arriadne; print(arriadne.__version__)"
# Copy plugin to ~/.hermes/plugins/ariadne, then:
hermes config set memory.provider ariadne
hermes memory statusHindsight
docker run -it --pull always --name hindsight --restart unless-stopped -p 8888:8888 -p 9999:9999 -e HINDSIGHT_API_LLM_API_KEY=$OPENAI_API_KEY -v hindsight-data:/home/hindsight/.pg0 ghcr.io/vectorize-io/hindsight:latest
curl http://localhost:8888/health
hermes memory setup
hermes memory statusConclusion
My recommendation: test Mnemosyne → Open Second Brain → Ariadne in that order. If those do not meet the bar, test Hindsight as the heavier service-backed contender. Keep agent-memory-skill and hermes-memlock in mind as companion plugins, not primary providers.