feat(eval): memory eval foundation — research survey + L1/L2/L3 harness + frozen baselines - #2
Merged
Merged
Conversation
… 50-100 QA cases)
…lement never receives test writes
…ram enablement/config
…l ~/.engram enablement/config" This reverts commit d019221.
…ake enablement never receives test writes" This reverts commit 91ef1ba.
…judge failures, checkpoint/resume Judge calls now retry up to 3 attempts (10s/30s backoff), capture claude's stderr on ExitError (truncated to 500B), and on final failure return a sentinel verdict (Score -1) instead of fataling the whole grid. runL3 excludes judge-failed items (Score < 0) from mean_score_<arm> rather than counting them as 0, marks them with per-item judge_failed:1, emits judge_failed_count, and warns loudly to stderr when an arm exceeds 20% failures. New eval/e2e/checkpoint.go writes an incremental <out>.runs.jsonl sidecar (run line before judging, verdict line after, kind-discriminated, flushed per line) and replays it on startup to skip completed work and re-judge-only work. Pure serialize/parse/classify logic is unit-tested (no claude calls).
…laims to match shipped behavior
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #3
Summary
Phase 0 + Phase 1 of the memory eval & optimization program (spec:
docs/superpowers/specs/2026-07-24-memory-eval-optimization-design.md, plan:docs/superpowers/plans/2026-07-24-memory-eval-foundation.md).Program goal: double the memory uplift and halve the memory-injected token cost of engram+MemoryLake, measured on phoenix as testbed. This PR delivers the research survey and the complete measurement foundation, with baselines frozen.
Phase 0 — Research
docs/research/2026-07-24-memory-sota-survey.md: 38 techniques across 4 directions, 29-row priority matrix feeding Phase 2.Phase 1 — Eval foundation (
eval/, never ships in the release binary)phoenix-retrieval-v1), recall@k/MRR/token/latency against live MemoryLake; LLM-verify fallback for keyword misses.mem_context+ retrieval composite.phoenix-e2e-v1), isolated no-memory vs memory arms for headless Claude Code, calibrated single-turn LLM judge (tools disabled), checkpoint/resume, judge retry.memorylake.ListAllFacts,mcp.ServerInstructions).Baseline anchors (Phase 2 acceptance)
Key insight: all uplift comes from gotcha tasks (+8.50); memory is currently a net drag on architecture-qa (−0.90) and small-fix (−0.80), and the memory arm emits +48% output tokens. Full analysis in
eval/results/baseline-report.md.Notes
5e44d66).🤖 Generated with Claude Code