atmem/code-discovery-benchmark-evidence-20261020

Repository discovery: native lexical, Cognee CODE and AtMem

Twelve real subscription-backed coding attempts on a historical Click issue, with executable grading, complete token receipts, tool traces and recoverable encrypted evidence. Two distinct conditions are provided: optional discovery and an explicit request to begin with a self-selected discovery query. Six metric rows from one SWE-bench Verified qualification task are also provided, separately identified in the same table. Its raw evaluator evidence stays private.

Explore the data

Select trials in the dataset viewer for the comparison table. Each row shows the condition, agent, backend, task outcome, total input including cache, output, agent/index/setup time, lookup attempts and nonempty lookup responses. Select traces to inspect readable event payloads and exact tool requests/results. The traces configuration covers Click and its qualification launches, not the private SWE evaluator. The native locator's transient AST preparation is included in setup/agent timing; the index column measures persistent graph construction.

Field	Meaning
protocol	Optional or requested-consultation condition; do not pool them
host, backend	Coding CLI and configured discovery backend
task_passed	Entire frozen executable grader passed
input_tokens_including_cache	Full session input, not just retrieved text
agent_seconds	Coding session wall time, excluding preparation/grading
index_seconds	Separately measured graph construction
discovery_attempts	Calls attempted, including malformed requests
nonempty_discovery_responses	Successful responses containing locations

All backends use the same navigation schema and offline source read/edit/test tool. The native comparator is an AST-backed lexical locator, not an untouched native CLI. Cognee uses its actual pinned deterministic CODE extraction/query path, not AtMem's incorporated implementation or its other search modes.

Evidence and interpretation

This is one previously inspected task, one attempt per cell per condition: a diagnostic, not a full coding-benchmark score. The optional condition had zero navigation use. An error or empty lookup is not treated as use of code evidence; token differences alone cannot establish a backend benefit.

archives/ contains original signed/encrypted event exports. Recovery keys stay with the owner and are not uploaded. Readable traces are decoded, privacy-redacted derivatives with separate hashes; original signatures do not apply to them. Independent recovery is checked with access to the original vault disabled. All qualification failures are retained separately from scored task attempts.

If present, swe-pilot-metrics.json contains only the metric projection of one SWE-bench Verified qualification task. It is not the full 500-task benchmark. SWE evaluator/reference material is excluded from this public export.

Versions and licenses

AtMem source-preview wheel SHA256: 77a67773e33adcfe2c8a60d5a1da29326f0bc2dfe27663fc5f14297d0e272d92. The exact wheel is retained under artifacts/ for reproduction. It is a private source-preview build with 2.3.8 metadata, not a new public package release. Cognee source revision: f154b32a9e544dac73e7ddf2010327cf9988edcd. Click parent: 8c1a0a7abbc1c36f70d1f65f3604acc46c5ce6ab. Coding agents: Claude Code/Opus 5.5 and Codex/GPT-6 Luna; host/model identities are retained in traces. Copilot's native account policy blocked third-party MCP, so it has no scored cells. No policy bypass or paid API fallback was used.

AtMem/Cognee code is Apache-2.0; Click code is BSD-3-Clause. See the included license files. These are evaluation records, not a memory-training corpus.