Modes Explained

Modes Explained

SuperLocalMemory V3 offers three operating modes. Choose based on your privacy requirements and accuracy needs.

Mode A: Local Guardian

Zero cloud. Maximum privacy.

• All memory operations run locally on your machine
• No API calls, no cloud services, no data transmission
• Retrieval uses 4-channel hybrid search: semantic similarity, keyword matching, entity graph traversal, and temporal relevance
• Cross-encoder reranking for precise result ordering
• Mathematical foundations (Fisher-Rao, Sheaf, Langevin) enhance accuracy without any LLM
• EU AI Act compliant by architecture — data never leaves your device

Who it's for: Privacy-conscious developers, enterprise environments with strict data policies, EU-regulated industries, air-gapped systems.

Limitations: No LLM-powered answer synthesis. Returns ranked memory excerpts rather than composed answers. Best accuracy on factual and entity-based queries.

Benchmark: 74.8% retrieval accuracy on LoCoMo (10 conversations, 1,276 questions). 85.0% on open-domain questions — the highest of any system evaluated. Outperforms Mem0 (64.2%) without a single API call.

Mode B: Smart Local

Local LLM for answer synthesis. Still fully private.

• Everything in Mode A, plus a local LLM via Ollama
• The LLM synthesizes retrieved memories into coherent answers
• All processing stays on your machine — nothing sent to the cloud
• Requires Ollama installed with a model (e.g., llama3.2, mistral, phi3)

Who it's for: Developers who want composed answers but need data to stay local. Teams that can run Ollama on their machines.

Requirements:

• Ollama installed
• At least one model pulled: ollama pull llama3.2
• 8GB+ RAM recommended for good model performance

Limitations: Answer quality depends on the local model's capabilities. Smaller models may produce less accurate synthesis.

Mode C: Full Power

Maximum accuracy. Cloud LLM for fact extraction and answer synthesis.

• Everything in Mode B, plus cloud LLM support
• LLM-powered fact extraction for richer ingestion
• Agentic retrieval with multi-round refinement
• Supports Azure OpenAI, OpenAI, Anthropic, and other providers

Who it's for: Developers who prioritize accuracy over privacy. Teams with approved cloud AI policies. Research and benchmarking.

Benchmark: 87.7% on LoCoMo conv-30 (81 questions). 100% on multi-hop questions. Competitive with funded systems like Zep v3 (85.2%) and approaching EverMemOS (92.3%).

Note: Data is sent to the cloud provider you configure. Ensure your organization's policies allow this.

Switching Modes

Check your current mode:

slm mode

Switch modes:

slm mode a    # Switch to Local Guardian
slm mode b    # Switch to Smart Local
slm mode c    # Switch to Full Power

Mode changes take effect immediately. Your stored memories are not affected — all modes use the same database.

Comparison Table

Feature	Mode A	Mode B	Mode C
Semantic search	Yes	Yes	Yes
Keyword search (BM25)	Yes	Yes	Yes
Entity graph	Yes	Yes	Yes
Temporal retrieval	Yes	Yes	Yes
Mathematical scoring	Yes	Yes	Yes
Cross-encoder reranking	Yes	Yes	Yes
LLM fact extraction	No	Local	Cloud
LLM answer synthesis	No	Local	Cloud
Agentic retrieval	No	No	Yes
Data leaves device	Never	Never	Yes
EU AI Act compliant	Yes	Yes	Partial
Internet required	No	No	Yes

Recommendations

• Start with Mode A if you are unsure. You can always upgrade later.
• Use Mode B if you have a capable machine (16GB+ RAM) and want composed answers locally.
• Use Mode C for maximum accuracy when cloud access is acceptable.

---

Part of Qualixar | Created by Varun Pratap Bhardwaj