01
96.6%Top-1 accuracy
Target file ranked first.
Public benchmarks
High-signal code context.Fewer tokens spent finding it.
Memtrace returns precise structural context instead of broad code chunks. In the public exact-symbol run, it returned roughly 90% fewer response tokens than ChromaDB while keeping top-1 accuracy above 96%.
1,000 exact-symbol queriesPublic inputs and outputs
TOKEN ECONOMYCorrect answers per token
1K queries495.52
correct top-1 hits per 1K response tokens
3.9×
the next closest system on correct top-1 hits per response token
Exact-symbol retrieval
The right code appears first.
A coding agent saves little when compact context points to the wrong file. Memtrace keeps the result precise while returning far less text.
02
96.7%Precision@10
Penalizes extra paths around a correct hit.
03
99.7%Recall@10
Target file found in the first 10.
04
0.979MRR
Scores an earlier first correct result higher.
WORKLOADmempalace exact-symbol retrieval1,000 queries
Context efficiency
More answer.Less context to process.
Broad retrieval fills the context window with nearby code. Memtrace returns the structural path the agent asked for, leaving more room for the change itself.
84.7%fewer file-content tokens than raw reads
Typical session benchmark using compressed source readsTOTAL RESPONSE TOKENSLOWER IS BETTER
90%
fewer response tokens than ChromaDB on the same 1,000 queries
Runtime performance
Fast enough to check before every edit.
Context only helps when it arrives inside the coding loop. These measurements cover graph lookup, freshness, indexing and memory use on the public Django and incremental fixtures.
0.19 ms
average graph lookup
Django · 50K-node graph
42.5 ms
p95 update-to-query time
50-edit incremental fixture
13.6 sec
repository to queryable graph
Django · 3.3K files
26.2 MB
peak process memory
isolated Django process
Methodology
Public inputs.Raw results your team can inspect.
Versions, inputs, raw outputs, and scoring stay public. Your team can rerun the benchmark and inspect every result.
Run the same query set.
Every system receives the same 1,000 exact-symbol questions against the same repository.
Score the returned ranking.
Accuracy, precision, recall and MRR come directly from where the correct file appears.
Inspect the raw output.
The inputs, result files and scoring method are public, so your team can run the test again.
Put the context to work
Give every agent the code context it needsbefore the next edit.
Start with one repository. Add team memory and coordination when your rollout grows.