Research
Benchmarking a Token-Efficient Memory Algorithm for AI Agents
Unispec's token-efficient memory algorithm achieves high accuracy on LoCoMo, LongMemEval, and BEAM while using a fraction of the tokens of full-context approaches.
Benchmark results
Latency measured on the hosted instance (n=20/op, 2026-07-03). We publish only figures we can reproduce — quality benchmarks will follow.
Memory API (hosted, measured)
- Add (infer=false) p50
- 78 ms
- Add (infer=false) p95
- 253 ms
- Semantic search p50
- 72 ms
- Semantic search p95
- 99 ms
Vector Database (hosted, measured)
- Text query p50
- 88 ms
- Text query p95
- 100 ms
- Upsert-text p50
- 76 ms
- Upsert-text p95
- 189 ms
Quality benchmarks
- LoCoMo / LongMemEval / BEAM
- publishing after reproduction
What's New
Two advances under the hood
Single-Pass ADD-Only Extraction
Treat agent-generated facts as first-class, extracted in a single pass without redundant re-processing.
Multi-Signal Retrieval
Three scoring passes run in parallel — semantic similarity, keyword matching, and entity matching — then fused for relevance.
What We're Building Next
The road ahead
Temporal abstraction
Reason about when facts were true and how they evolve over time.
Cross-session structure
Connect memories across sessions into coherent, queryable structure.
Agent-native memory
Memory primitives designed for how agents actually plan and act.