Skip to content

Performance Benchmarks ​

LimbicDB is designed to handle thousands of memories efficiently within AI agent environments. The following benchmarks represent a baseline for the current implementation, demonstrating performance across different search mechanisms.

Generated: April 2026

Hardware Context

Results are environment-specific and should be interpreted as relative comparisons rather than absolute limits. They were compiled on a standard developer machine (macOS/Windows, typical multi-core CPU).

When using persistent storage via ./agent.limbic (the default mode), memory retrieval is exceptionally fast for keyword searches, and scales linearly for semantic searches.

MemoriesKeyword (ms)Semantic (ms)Hybrid (ms)Notes
1000.22.62.0-
10000.421.018.4-
50001.5113.095.0-

Memory Backend (Ephemeral) ​

The pure :memory: backend avoids all disk I/O, providing instantaneous keyword lookups. Currently, it does not compute vector embeddings by default.

MemoriesKeyword (ms)Semantic (ms)Hybrid (ms)Notes
1000.2N/AN/ANo embeddings;
10000.2N/AN/ANo embeddings;
50001.0N/AN/ANo embeddings;

Interpretation & Next Steps ​

  • Keyword Search: Leverages SQLite's native FTS5 engine or fast in-memory matching. Extremely performant and suitable for large datasets.
  • Semantic Search: Utilizes exact cosine similarity matching. While this guarantees 100% accurate recall against the provided vectors, the O(N) complexity causes latency to scale linearly.
  • Hybrid Search: Merges keyword indexing speed (30% weight) and semantic accuracy (70% weight), carrying a similar computation cost to pure semantic search.

Future Optimizations ​

As LimbicDB evolves beyond the alpha stage, the following performance enhancements are planned:

  • HNSW (Hierarchical Navigable Small World) Indexing: To transition semantic search scaling from O(N) to O(logN) using vector quantization extensions like sqlite-vss.
  • WebAssembly Backends: Improving cross-platform embedding computation speed within edge environments.

Released under the MIT License.