v3.5.0 Scale-Ready — CozoDB + LanceDB

1 Million Memories.
1-Second Recall.

SuperLocalMemory v3.5.0 delivers the full 3.4.5 "Scale-Ready" architecture into production. CozoDB and LanceDB are now on the recall path. All 6 channels contribute. Recall completes in under 1 second. Your database auto-migrates on upgrade.

pip install -U superlocalmemory && slm restart

Get Started

Performance

Real measurements on a production database with 17,500+ facts and 380K graph edges.

<1s

warm recall (all 6 channels)

down from 13-19s

20ms

BM25 channel (FTS5)

was 11.2s (rank_bm25)

0.4s

Hopfield channel (ANN prefilter)

was 6s (full-matrix)

100%

vector store coverage

auto-backfill on startup

What's New in v3.5.0

CozoDB + LanceDB on Recall Path

The 3.4.5 pluggable backend architecture is now wired into every recall. Auto-detected, auto-migrated.

→ CozoDB: entity_graph channel routes through SSI-indexed graph
→ LanceDB: vector store with ANN indexing
→ Auto-migrate: background thread on daemon start
→ SQLite fallback if backends unavailable

6-Channel Recall <1s

Every channel tuned for speed. Consistent sub-second recall across all surfaces: MCP, CLI, Dashboard.

● BM25: rank_bm25 11.2s → FTS5 20ms (C-level index)
● Hopfield: full-matrix 6s → ANN prefilter 0.4s
● Temporal: dateutil 2.6s → fromisoformat 0.25s
● Scene expand: 20 individual queries → batch scan

Core Memory Block

Always-injected high-value facts at the top of every session — the agent never forgets the essentials.

★ Auto-derived: importance ≥ 0.8 or access_count ≥ threshold
★ Explicit pin via core_memory MCP tool (pin/unpin/list)
★ Migration M015 — additive pinned column, backward-compatible
★ Pinned facts surface even when the query doesn't match

Context Injection v2

Unified formatter across all surfaces — MCP, CLI, and Dashboard produce identical injection.

• Token-budgeted injection (mode-aware: 2K/4K/8K tokens)
• Full-fidelity content — no more 200-char stubs
• Edge-placement ordering (strongest at top + bottom)
• Configurable trust_first_party (false shipped, true personal)

Score Normalization [0, 1]

All fusion scores mapped to a clean 0–1 range via soft sigmoid. Readable. Verifiable. Rank-order preserved.

● Soft sigmoid applies after ranking pipeline (v1/v2/bandit)
● Same normalization on MCP, CLI, and Dashboard
● Scores no longer unbounded (was 27+, now [0, 1])

Vector Store + Quality Gate

Auto-backfill indexes every fact with embeddings. Store quality gate rejects template/placeholder junk.

● Vector store backfill from 5.8K → 17.4K (100% coverage)
● Store gate: is_low_quality() drops template leaks before insert
● Pending materializer: 50/sec drain (was 2.5/sec)

How to Upgrade

One command. Your database migrates automatically — including CozoDB, LanceDB, and the pinned column.

pip install -U superlocalmemory

slm restart

✓

Schema v3.5.0 applied. CozoDB + LanceDB detected and migrated. Vector store backfilled. Core Memory Block active.

No manual commands. No data loss. Zero downtime. Your database upgrades in-place on daemon restart. CozoDB and LanceDB are auto-detected — just install the packages.

The Architecture — v3.5.0

SQLite stays canonical. CozoDB and LanceDB on the recall path — migrated, wired, and delivering sub-second recall.

┌──────────────┐  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐
│   SQLite     │  │   CozoDB     │  │   LanceDB    │  │  FTS5 Index  │
│  (CANONICAL) │  │  (GRAPH)     │  │  (VECTOR)    │  │  (BM25)      │
│              │  │              │  │              │  │              │
│ 17.5K facts  │  │ 380K edges   │  │ 17.4K vecs   │  │  20ms MATCH  │
│ 1.7GB        │  │ 31ms SA      │  │  364ms hopf  │  │  C-level     │
└──────┬───────┘  └──────┬───────┘  └──────┬───────┘  └──────────────┘
       │                 │                  │
       └──── ALL ON RECALL PATH ────────────┘
                    │
              ┌─────▼──────┐
              │ [MCP/CLI/Dashboard] │
              │ 6-channel <1s        │
              │ Core Memory + Pinned  │
              └────────────────────┘

Ready to Scale?

Install once. Every session remembers the last. Automatically. Forever. v3.5.0 — CozoDB. LanceDB. Sub-second recall.

Get Started GitHub

1 Million Memories. 1-Second Recall.