Skip to content

Home

Jump to bottom
Varun Pratap Bhardwaj edited this page Jul 16, 2026 · 39 revisions

SuperLocalMemory V3.7

Local-first agent memory, retrieval, cache, compression, and trusted-peer coordination in one operator-controlled runtime.

SuperLocalMemory turns conversations, observations, and connected-source evidence into durable memory that can be recalled through a CLI, MCP, hooks, dashboard, or documented IDE integrations. SQLite + sqlite-vec are the canonical local store. The product also includes an explicit Scale Engine for CozoDB graph and LanceDB vector projections, a cache/compression module, and SLM Mesh coordination controls.

The product in one view

Sources and clients
CLI · MCP HTTP/stdio · hooks · dashboard · IDEs · adapters
                              │
                              ▼
  admission → queryable core → enrichment → brain/lifecycle
                              │
                              ▼
 semantic · BM25 · temporal · Hopfield · spreading activation
                              │
                              ▼
 safe bounded context with policy, provenance and trace evidence
                              │
                              ▼
 SQLite + sqlite-vec canonical ─► parity-gated graph/vector projections

The architecture has seven logical stages: admission, queryable durability, enrichment, learning/lifecycle, retrieval, safe context delivery, and operations. A specific write or recall only reports stages that actually ran; optional enrichers and retrieval channels are dependency- and mode-aware.

Capability map

Area Available capability Important boundary
Memory Facts, scenes, temporal events, entities, profiles/scopes, memory lifecycle Recalled content is untrusted evidence, never a new instruction.
Ingestion Replay-safe operation receipts; extraction, entity, graph, temporal, provenance and embedding derivations Use --sync when a caller needs all declared stages, not only the immediate queryable receipt.
Recall Semantic, BM25, temporal, Hopfield and spreading-activation candidates; fusion, optional rerank and graph score enhancement Runtime health determines the channels that participate.
Brain Behavioral patterns, feedback/outcomes, reward signals, consolidation, soft prompts and guarded skill evolution Learning is not a guarantee that an outcome was correct or beneficial.
Graph Canonical entities, aliases, profiles, edges, scenes, timelines and an Entity Explorer Graph evidence is inspectable and provenance-bearing.
Scale Engine CozoDB graph + LanceDB vectors with prepare → verify → promote → rollback SQLite remains canonical; promotion is explicit and parity-gated.
Optimize Exact cache, tag invalidation, safe compression, opt-in lossy prose compression and CCR originals Only the proxy can intercept a primary provider turn.
Mesh Authenticated peer messages, locks, inbox/outbox, queues and optional discovery Mesh coordinates peers; it is not a replicated distributed-memory database.
Governance Provenance, audit, retention, policy, export/erasure, health and diagnostics Deployment configuration determines compliance posture.
Integrations CLI, Python SDK, MCP, Claude plugin, Codex add-on, documented IDE configs, Gmail/Calendar/transcript adapters Connectors and hooks are opt-in and have their own data paths.

Operating modes

Mode Core behavior Model path
A — Local Guardian Local core memory and math-informed retrieval No cloud model provider is required for core operations.
B — Smart Local Mode A plus an operator-managed Ollama endpoint Local LLM endpoint.
C — Provider-assisted Local storage with configured provider-backed enrichment/retrieval behavior Content sent to the configured provider follows that provider path.

Mode A does not disable model downloads, adapters, backup, proxy providers, or other integrations that an operator explicitly enables. Review the complete deployment before making a privacy or compliance determination.

Dashboard workspaces

The local dashboard includes Dashboard, Brain, Knowledge Graph, Memories, Health, Operations, Entity Explorer, Skill Evolution, Mesh Peers, Settings, and Optimize workspaces. Use it with slm health, slm doctor, and slm trace for operational verification rather than treating a visual status as a guarantee.

Quick Start

npm install -g superlocalmemory    # Primary global CLI path
slm setup                          # Choose mode A/B/C
slm warmup                         # Pre-download embedding model (optional)

The second primary path is Python in an activated virtual environment:

python3 -m venv .venv
source .venv/bin/activate  # Windows PowerShell: .venv\Scripts\Activate.ps1
python -m pip install superlocalmemory
slm setup

Then configure the client you intend to use and verify it with slm doctor.

SuperLocalMemory V3 — Local-first memory with explicit data-path controls.

Part of Qualixar | Created by Varun Pratap Bhardwaj | GitHub

SuperLocalMemory V3

Getting Started

Reference

Architecture

Enterprise

V2 Documentation

Clone this wiki locally