SuperLocalMemory Logo — Local AI Memory Layer
SuperLocalMemory
v3.4.5 Scale-Ready

1 Million Memories.
Zero Slowdown.

SuperLocalMemory now scales to five years of daily AI use without performance degradation. Tiered storage, graph pruning, and optional acceleration backends. Tested on 1.18 million real graph edges.

pip install -U superlocalmemory && slm restart
Get Started

What's New in v3.4.5

Tiered Memory Lifecycle

Every memory auto-classifies as active, warm, cold, or archived by age and access frequency.

  • ● Active: hot path retrieval, full graph + vector
  • ● Warm: reduced weight, still searchable
  • ● Cold: on explicit deep recall only
  • ● Archived: compressed, never deleted

Graph Pruning Engine

Three strategies keep your graph lean without losing semantic meaning.

  • ✓ Chain collapse: A→B→C → A→C when direct path is stronger
  • ✓ Garbage entity removal
  • ✓ Low-activity edge decay (90+ days)

Pluggable Backend Architecture

Optional acceleration for graph and vector operations. Zero-config, auto-detect.

  • → Graph: CozoDB for PageRank at 1M+ edges
  • → Vectors: LanceDB with IVF+PQ indexing
  • → Falls back to SQLite+NetworkX if not installed

Applies to Existing Data

Migration is fully automatic. No `slm migrate` needed.

  • • Schema applied via M014 on daemon restart
  • • Existing facts auto-tiered by age
  • • Backends initialize in background threads

No Data Loss. Ever.

SQLite stays the canonical source of truth at all times.

  • ◆ ALTER TABLE ADD COLUMN — non-destructive
  • ◆ Backends are derived views, rebuilt from SQLite
  • ◆ Rollback: pip uninstall pycozo → falls back to NetworkX

Backup Frequency Reduced

Backups now run every 7 days instead of daily. Only 2 backups retained.

  • ↓ From 8.2GB backups to <2GB
  • ↓ From 60 backup files to 2

How to Upgrade

One command. Your database migrates automatically.

$
pip install -U superlocalmemory
slm restart
✓
Schema v3.4.5 applied. Backends detected. Ready.

No manual commands. No data loss. Zero downtime. Your database upgrades in-place with no user action required beyond the restart.

Before & After

Tested on a production database with 1.18M graph edges and 16K facts.

<1ms
health endpoint
was 16,000ms
356MB
daemon RAM
was 1.2GB
~2s
recall latency
was timed out
1.18M
graph edges
handled

The Architecture

SQLite stays canonical. CozoDB and LanceDB are derived views — rebuildable on demand.

┌──────────────┐  ┌──────────────┐  ┌──────────────┐
│   SQLite     │  │   CozoDB     │  │   LanceDB    │
│  (CANONICAL) │  │  (DERIVED)   │  │  (DERIVED)   │
│              │  │              │  │              │
│ 16K facts    │  │ 200K hot     │  │ hot+warm     │
│ 1.18M edges  │  │ edges only   │  │ vectors only │
│ 850MB        │  │ PageRank     │  │ cosine ANN   │
└──────┬───────┘  └──────┬───────┘  └──────┬───────┘
       │                 │                  │
       └──── ALWAYS ─────┴─── WRITTEN ──────┘

Ready to Scale?

Install once. Every session remembers the last. Automatically. Forever.

Get Started GitHub