# memoryleaderboard

Entry for demarkus in the Agent Memory Leaderboard: a Go adapter exposing the benchmark's Add/Search HTTP contract over a per-user demarkus world, with agentic search as the submission.

## Status

The repository is public at [latebit-io/memoryleaderboard](https://github.com/latebit-io/memoryleaderboard). Provider diagnostics, bounded 429 retry, empty-result handling, and evidence-selection tuning merged through PR #6; `main` is now `874c86f`. Live `gpt-4o-mini` validation is complete: smoke passes with zero noise; provider-backed quality has perfect recall, evidence recall, nDCG, and MRR with one extra rejected proposal; conformance is 18 pass/0 fail; and a 16-worker, 32-user capacity run completed 32 Adds plus 64 Searches with zero HTTP, schema, isolation, or visibility failures. Search p95 was 6.57 seconds. Repeated-memory Top-K runs retained a 5-7/64 exhaustive-visibility gap despite otherwise clean operation. Full race tests, vet, shell checks, and diff checks pass. Next gate is host deployment.

## Plan

Authoritative plan: [Agent Memory Leaderboard entry](/plans/agent-memory-leaderboard.md). Search uses a scoped nib navigation agent over Demarkus lookup/list/fetch; Add uses `gpt-4o-mini` distillation while retaining raw evidence. Production is a single encrypted Linux host with Caddy TLS, adapter, private Demarkus storage, and explicit 30-day data/backup deletion. The Academic submission uses Coding Memory and the hosted Add/Search API route.

Echo v5 is the Go HTTP framework used by the adapter, not an LLM provider.

## Phases

| # | Work | Output |
|---|------|--------|
| 0 | Recon: challenge repo, limits, cycle dates, route | Academic hosted API route selected |
| 1 | Adapter skeleton: Add/Search against local Demarkus | merged |
| 2 | Agentic search; catalog fallback | merged via PR #2 |
| 3 | Ingestion tuning: distillation, temporal metadata, idempotency | merged via PR #3 |
| 4 | Conformance, synthetic retrieval quality, Top-K load evaluation | live provider gates complete; repeated-memory Top-K gap recorded |
| 5 | Public release, provider validation, host deploy, Smoke, Full | public release and provider validation complete; host deploy pending |

## Repo

[github.com/latebit-io/memoryleaderboard](https://github.com/latebit-io/memoryleaderboard) is public. Local checkout: `/Users/fritz/latebit/memoryleaderboard`, clean `main` at `874c86f` matching `origin/main`.

## Workspace

- [Agent Memory Leaderboard](/memoryleaderboard/agent-memory-leaderboard.md) official benchmark bookmark and entry links
- [Agent Memory Leaderboard entry plan](/plans/agent-memory-leaderboard.md) active roadmap
- [journal/](/memoryleaderboard/journal/) session notes
- [debugging.md](/memoryleaderboard/debugging.md) investigation lessons and portability gotchas
- Add architecture.md and patterns.md as they earn content.