FAQ

FAQ — V4.0.3

Frequently asked questions about SuperLocalMemory V4.0.3.

General

What is SuperLocalMemory?

SuperLocalMemory is a persistent memory system for AI assistants. It stores your decisions, bug fixes, project context, and preferences locally, then automatically provides them to your AI in future sessions via 5-channel retrieval (semantic, BM25, temporal, Hopfield, spreading activation — plus graph enhancement). Your AI stops forgetting you.

Is it really free?

Yes. SuperLocalMemory is open-source (GNU Affero General Public License v3.0 or later) and completely free. No usage limits, no credit system, no subscription. Forever.

Where is my data stored?

Core memory is SQLite-backed inside the configured SLM data root. That root also contains configuration, logs, queues, models, and derived state. Mode C sends configured query or enrichment content to its provider; optional connectors, backup, and downloads have their own network behavior in every mode.

Which IDEs and platforms are supported?

Run slm connect --list for the release's documented client names. MCP-compatible clients can also be configured manually, but a client is considered verified only when it passes the release integration matrix. V4 platform boundary: Apple Silicon macOS, 64-bit Windows, 64-bit Linux — Intel Mac and 32-bit Windows are not supported (cryptography==50.0.0).

Does it work offline?

Mode A and Mode B work fully offline. Mode C requires internet for the cloud LLM. Optional connectors, backup, and model downloads require network in any mode when explicitly enabled.

Installation

What are the requirements?

• Python 3.11 – 3.14 (required for V4 engine)
• Node.js 18+ (if installing via npm)
• Platform: Apple Silicon macOS, 64-bit Windows, or 64-bit Linux (Intel Mac / Win32 not supported)
• Any supported IDE
• For Mode B: Ollama with a pulled model
• For Mode C: API key for your cloud LLM provider

How do I install it?

# npm (recommended)
npm install -g superlocalmemory
slm setup
slm warmup    # Optional — pre-download embedding model

# or inside an activated Python virtual environment
python3 -m venv .venv
source .venv/bin/activate  # Windows PowerShell: .venv\Scripts\Activate.ps1
python -m pip install superlocalmemory
slm setup

See Installation for platform notes and troubleshooting.

How do I update?

npm install -g superlocalmemory@latest
# or, while the SLM virtual environment is active:
python -m pip install --upgrade superlocalmemory
# then:
slm restart && slm doctor

I am upgrading from V2. Will I lose my data?

No. Run slm migrate (V2→V3 data migration) after updating — or slm db migrate for additive schema maintenance; they are different commands (see CLI Reference). All memories, profiles, and settings are preserved. A backup is created automatically. See Migration from V2 for details.

Usage

How does auto-recall work?

When you start a conversation in your IDE, SuperLocalMemory automatically retrieves relevant memories and injects them into your AI's context via the MCP server (slm mcp or HTTP http://127.0.0.1:8765/mcp/). You do not need to call "recall" explicitly — it happens according to the client's configured instructions/hooks.

How do I store a memory?

slm remember "The deploy script needs AWS_REGION set to us-east-1"
slm remember "Decision" --scope shared --shared-with team-a

What do queryable, enriching, complete, and failed mean?

• queryable means raw evidence and the SQLite relational/FTS projection are durable and recallable.
• enriching means a lease-owning worker is running configured derivation stages.
• complete means every declared derivation and configured projector succeeded (and a hash-verifiable manifest is sealed).
• failed retains the raw evidence, error, attempt count, and retry timing; it is not silent data loss (see slm ops list / list_failed_operations).

How do I search memories?

slm recall "deploy configuration"

Current recall uses five candidate producers (semantic, BM25, temporal, Hopfield, spreading activation) plus entity-graph score enhancement.

How do I see which retrieval channels found what?

slm trace "deploy configuration"

This shows per-channel scores (Semantic, BM25, Temporal, Hopfield, Spreading Activation) for each result. Entity-graph data can enhance a post-fusion score but is not a separate recall channel. The current implementation is five producers, not four — earlier docs that said four-channel are obsolete.

How do I delete a memory?

slm forget "search query"     # Delete matching memories (with confirmation)
slm delete <fact_id> --yes    # Delete one fact by ID (use slm list to find IDs)

Use slm ops list or list_failed_operations (MCP, power/whole profile) to inspect stuck operations.

Modes

Which mode should I use?

• Mode A if you need privacy, compliance, or offline operation
• Mode B if you want composed answers and have a capable machine (16GB+ RAM)
• Mode C if you want maximum accuracy and cloud access is acceptable

Can I switch modes after setup?

Yes: slm mode a, slm mode b, or slm mode c. Your memories are shared across all modes.

What are the accuracy differences?

The V3 paper provides published LoCoMo evidence carried into V4 (not a newly rerun V4 benchmark) — arXiv:2603.14588:

• 60.4% Mode A Raw across 10 conversations / 1,276 questions with zero-LLM answer construction.
• 74.8% Mode A Retrieval across the same scope with local retrieval and GPT-4.1-mini answer synthesis.
• 87.7% Mode C on Conv-30 / 81 questions with cloud embeddings and GPT-4.1-mini answer generation and judge.

The figures retain their original protocol scope; they are not a newly rerun V4 package benchmark and are not comparable across vendors without matching protocol (conversation scope, question count, retrieval stack, answer model, judge, release artifact). See the linked preprint for category results, ablations, and limitations. The non-rerun nature is explicit in README and docs/benchmarks.md.

V4 Reliability Contract

What is actually verified with 2,200/2,200?

Only one stress figure is verified in V4.0.0: 2,200/2,200 trials (100%) from benchmark/run_all.py --trials 200 (11 experiments × 200). Source: benchmark/results/SUMMARY.md (4.0.0, Python 3.13.13, macOS-26.5.2-arm64, 2026-08-08) and benchmark/README.md honesty notes (exp1 embedding_metadata-only without sqlite-vec ANN, exp2 lightweight _TrackingOwner, exp7 _generation set directly, etc.). No universal latency, p99, or throughput claim is made — use the measured exp_governed_latency p50 or run the harness on your own machine.

Privacy and Security

Can anyone else see my memories?

No. Your database is a local file on your machine. It is not synced, uploaded, or shared with anyone — including us — unless you explicitly enable Mesh peering, cloud backup, or a provider-backed mode.

Does it guarantee regulatory compliance?

No software package certifies the complete deployment. SLM supplies local storage, memory erasure (slm forget/slm delete) and dashboard profile erasure, provenance, retention, access-policy (RBAC — see RBAC and Teams), and hash-chained audit controls (see GDPR Compliance and Compliance); applicability and sufficiency depend on the operator, use case, configuration, providers, and surrounding systems.

Can I export my data?

The database is a standard SQLite file at ~/.superlocalmemory/memory.db. Use slm evidence export for a checksummed JSONL bundle (see slm evidence --help and GDPR Compliance), or copy the data root. Configuration, logs, queues, models, derived indexes, and optional backend state also live in the data root.

Can I delete all my data?

slm forget "query" deletes matching memories, and slm delete <fact_id> deletes an exact fact. V4 does not expose a slm profile delete CLI command; non-default profiles can be erased through Dashboard → Governance → Data Privacy → Erase with typed confirmation. To delete the complete installation, follow the documented erasure/uninstall procedure — do not assume removing only memory.db covers configuration, logs, queues, models, derived indexes, and optional backend state.

Troubleshooting

My AI does not seem to remember anything.

1. Check that SuperLocalMemory is running: slm status / slm health
2. Check that you have stored memories: slm recall "test" --json
3. Verify your IDE connection: restart the IDE after configuring MCP (slm connect --list)
4. Check the active profile: slm profile list
5. Inspect stuck ops: slm ops list or dashboard Operations

Recall returns irrelevant results.

Try more specific queries. Use slm trace "query" to see which channels contribute — this helps diagnose whether the issue is semantic, keyword, temporal, associative, or entity matching.

The setup wizard does not detect my IDE.

Use manual configuration. See IDE Setup for per-IDE config paths. Supported list via slm connect --list.

Where do I find Bounded Loops, Framework Adapters, GDPR/RBAC, Multi-Agent Memory?

• Bounded Loops: Bounded Loops + slm loop --help + MCP slm_loop_* (code/full/power/whole)
• Framework Adapters: Framework Adapters — 9 adapters under ide/integrations/
• GDPR/RBAC: GDPR Compliance, RBAC and Teams, Compliance
• Multi-Agent Memory: Multi-Agent Memory + SLM_AGENT_ID

Where can I report bugs?

Open an issue at github.com/qualixar/superlocalmemory/issues.

---

Part of Qualixar | Created by Varun Pratap Bhardwaj