Skip to content

Performance Benchmarks

Varun Pratap Bhardwaj edited this page Mar 16, 2026 · 1 revision

Performance Benchmarks

Measured performance of SuperLocalMemory on real hardware. Results reflect real-world use cases.


Search Latency

How fast SuperLocalMemory finds your memories.

Database Size Performance
Under 500 memories Sub-100ms search — faster than you can blink
Around 1,000 memories Sub-200ms — fully interactive response
5,000 memories Seconds range — optional indexes provide acceleration

What this means: For typical personal use (under 500 memories), search is effectively instant. Beyond 1,000 memories, optional acceleration indexes are available.


Concurrent Write Throughput

Multiple AI tools writing to memory simultaneously — the "database locked" problem, solved.

Scenario Result
1-2 agents writing simultaneously High throughput, sub-15ms latency, zero errors
5 agents writing simultaneously Moderate throughput, low latency, zero errors
10 agents writing simultaneously Stable throughput, zero errors

What this means: Zero "database is locked" errors, even with 10 AI tools writing at the same time.


Storage Efficiency

How much disk space your memories use.

Scale Approximate Size
1,000 memories ~1.5 MB
10,000 memories ~14 MB

What this means: Your entire AI memory history takes less space than a single high-res photo.


Knowledge Graph Construction

Building the relationship graph from your memories.

Scale Build Time
Under 100 memories Under 1 second
Around 1,000 memories Several seconds
5,000 memories Several minutes

What this means: Graph builds quickly for most users. The system consistently discovers natural topic communities across your memories. At 5,000 memories (the design limit), a full rebuild is an explicit design choice balancing graph utility against compute cost.

Graph Scaling: Knowledge graph features work best with up to 10,000 memories. For larger databases, the system uses intelligent sampling (most recent + highest importance memories) for graph construction. Core search and memory storage have no upper limit.


Trust Scoring — Memory Poisoning Defense

Bayesian trust scoring detects malicious agents attempting to corrupt your memory.

The trust system achieves strong separation between honest and malicious agents. Even a sophisticated "sleeper" attack — where an agent behaves well to build trust, then turns hostile — is detected with a significant trust drop. Zero false positives on benign agents.


Layer Contribution Analysis

The core retrieval system achieves high precision — the first relevant result is at position 1 for the vast majority of queries. The graph and pattern layers provide structural enrichment (clustering, relationship navigation, coding preferences) rather than modifying search ranking directly.


Coming Soon: LoCoMo Benchmark

We are currently running the LoCoMo benchmark (Snap Research, ACL 2024) — a standardized evaluation for long-conversation memory systems with multi-hop, temporal, and adversarial question types. Results will be published here when complete.


Methodology

  • All benchmarks run on local hardware with no cloud dependencies
  • Each measurement repeated multiple times with statistical aggregation
  • Database populated with realistic synthetic memories across diverse topics
  • Tests run on clean database instances for each benchmark scenario

See also: Architecture-V2.5 | Home

Clone this wiki locally