v1.7 — Submission freeze

The single frozen submission for Qwen Cloud Hackathon Track 1: MemoryAgent.
Exact SHA: abccc46 (policy/engine code state: 7dc465f; CI green on every commit in between).

ENGRAM is a verifiable memory control plane for Qwen agents. RAG retrieves documents; ENGRAM governs an agent's evolving beliefs and actions.

What this build proves, all on the frozen code (docs/evaluation.md, exact-SHA + UTC):

• Benchmark vs no-memory and full-history baselines: 5/5 (182 vs 453 tk prompt · zero stale recall · 48 tk store vs 1155 tk raw history · 3→1 consolidation · rescue at semantic 0.31)
• Ablations: semantic-only misses the allergy; no-arbiter serves two contradictory employer facts
• Action-aware server-side policy gate: 20/20 deterministic cases — deny precision 100%, false-block 0%; verdicts (deny / require-approval / preconditions) computed before generation, injected into the prompt, carried with a dry-run proposed_tool_call + dispatch{executed:false}, persisted in the per-turn audit
• Stability: belief revision 10/10 · critical rescue 10/10 (semantic 0.25–0.31)
• 14-spec Playwright E2E across desktop / 1024px / mobile — sessions & evidence live in drawers, never hidden

Final demo video (2:07): https://youtu.be/teJQ3MEEFJY — subtitle band, cursor/click/highlight annotations, policy chapter shows the server verdict AND its audit persistence.

Live: https://engram.hackthon.site · ?judge=1 · ?seed=devops
JUDGING.md · proof-of-deployment · evaluation

Post-freeze policy: uptime and security fixes only.

v1.6 — Final submission: Workbench + server-side policy gate

Final Qwen Cloud Hackathon Track 1 submission (exact SHA: 9013606, CI green on this commit).

Highlights:

• Conversation Workbench is the default product interface; Memory Lab isolates the constellation and Judge Demo
• Per-turn Memory Decision Audit (selected/rejected, score waterfall, injected context, ops — replayable from history, exportable as JSON)
• Server-side memory policy gate: a standing procedural rule denies a risky action BEFORE generation — verdict injected into the prompt, streamed to the UI, persisted in the turn audit
• Responsive evidence: below 1280px sessions/evidence become drawers (never display:none'd) — reachable on 1024px laptops and phones
• Streaming Stop, graph label pinning, store metadata (importance / created / used), workbench empty-state onboarding
• Atomic belief revision & consolidation, deferred reinforcement, cluster purity guard
• 14-spec Playwright E2E across 3 viewports; 10/10 belief-revision and 10/10 critical-recall stability; benchmark re-run 5/5 on final code
• Live: https://engram.hackthon.site · ?seed=devops · ?judge=1 · demo video (1:52): https://youtu.be/s0cigCj991U

JUDGING.md · proof-of-deployment · evaluation

v1.5 — Final submission: Conversation Workbench

The final frozen build for Qwen Cloud Hackathon Track 1: MemoryAgent.

The page is now organized around the user's current memory decision (reviewer round 6):

• Default mode is the three-pane Conversation Workbench: sessions | full-height conversation | live Memory Evidence (THIS TURN / GRAPH / STORE / DEMO tabs, COPY AUDIT JSON).
• Every answer carries a meta line — memories recalled · tokens · latency · ops · critical-rescue callout — that replays the frozen decision into the evidence pane.
• The console became MEMORY LAB (?mode=lab); the judge demo auto-opens it (?judge=1) and mirrors its 5/5 verdict back into the workbench.
• Memory policy gate: a standing procedural rule visibly blocks a risky DevOps action, with the rule and recall score shown before the agent acts.
• Visual denoise, conditional graph labels, prefers-reduced-motion, honest cross-run numbers (2.6–2.8×, 94–95%, semantic 0.25–0.31).

Verification on this build

• 11-spec Playwright E2E across desktop/laptop/mobile — all green in CI.
• Benchmark 5/5 re-run on final code; stability 10×: S2 10/10, S5 10/10 with latency p50/p95 in docs/evaluation.md.
• Live judge demo 5/5 from two external vantages (~50 s).

Final demo video (1:52): https://youtu.be/s0cigCj991U
Live: https://engram.hackthon.site · ?seed=devops · ?judge=1 · JUDGING.md

v1.4 — Final frozen submission

Frozen final for Qwen Cloud Hackathon Track 1: MemoryAgent (reviewer round 5 close-out).

Reliability

• Belief revision and sleep consolidation are atomic transactions; reinforcement commits only after a turn is delivered and persisted (commit_recall_usage); sleep clustering gains a centroid + pair-floor purity guard against union-find transitive bridges (hand-crafted-vector bridge test included).

Two real UI bugs found by the new E2E and fixed

• CSS smooth-scroll silently swallowed every programmatic scroll (streaming follow, pagination anchoring) — now inline-override instant.
• The dock resize grip's hit area was clipped to ~3px by overflow:hidden — now a real 12px strip.

Coverage

• Playwright across desktop / laptop (1024x768) / mobile (390x844): 10 specs incl. smart-follow scrolling, cursor pagination with viewport anchoring, drag-resize persistence, tab layout chat.

Consistency & verification

• All judge-facing links are the HTTPS domain (the app binds 127.0.0.1 behind nginx — documented, not exposed).
• Benchmark re-run on the final code state: 5/5 (2026-07-10, docs/evaluation.md).
• Live judge demo verified from two external vantages (CN direct / US exit): 5/5 in ~50 s each.

Live: https://engram.hackthon.site · mirror: https://engram.axiqo.xyz · DevOps: ?seed=devops
JUDGING.md · proof-of-deployment · evaluation

v1.3 — Memory Decision Console, hardened

Reviewer-driven hardening on top of the Memory Decision Console.

Security & correctness

• Session ownership enforced end-to-end: /api/messages, /api/chat and the new /api/turn_audit all 404 across users (asserted in tests/test_session_isolation.py over real HTTP; PRAGMA foreign_keys on).
• Sleep-cluster purity and belief revision covered by store-level tests.

Chat that works as a chat

• Drag-resizable / focusable chat dock (height persisted), smart follow-only-near-bottom streaming with a ↓ New-Response pill.
• Cursor pagination: full history reachable via Load-earlier (was capped at the last 40 messages).
• Per-turn memory-decision audits: every assistant turn freezes its selected + rejected candidates, score components, the exact context handed to Qwen, resulting memory ops, usage and latency — replayable from chat history into the Decision Inspector.

Tests & CI

• 4-spec Playwright E2E (console load, chat → audit replay, IME Enter guard, judge demo 5/5 live-verified) in offline fake-Qwen mode, plus a frontend JS syntax gate — all on every push.

Docs

• Judge demo timing, sleep threshold (.55 production default, env-tunable), benchmark cost and S5 similarity spread now match the code exactly.

Live: https://engram.hackthon.site · mirror https://engram.axiqo.xyz · DevOps scenario ?seed=devops
Judge quickstart: JUDGING.md · Proof: docs/proof-of-deployment.md

v1.2 — Final video release (Qwen Cloud Hackathon Track 1)

Final submission snapshot for Qwen Cloud Hackathon Track 1: MemoryAgent.

Demo video (2:42, narrated, native 1080p): https://youtu.be/yfjW6hGBj9s — cross-session critical recall → belief revision → sleep-cycle consolidation → DevOps runbook memory → benchmark/ablations → Alibaba Cloud proof.

Everything in v1.1 plus: re-recorded final demo video, embedded site favicon, architecture diagram refresh (permanent live-demo link).

• Live demo: https://engram.axiqo.xyz · DevOps scenario: https://engram.axiqo.xyz/?seed=devops
• Origin (Alibaba Cloud ECS): http://47.93.234.51:8080
• Judge quickstart: JUDGING.md
• Deployment proof: docs/proof-of-deployment.md
• Benchmark + ablations: docs/evaluation.md
• Offline check (no API key): ENGRAM_FAKE_QWEN=1 python3 tests/smoke_offline.py

v1.1 — Final judging release (Qwen Cloud Hackathon Track 1)

Final judging release for Qwen Cloud Hackathon Track 1.
Includes DevOps vertical demo, deployment proof, judge quickstart, ablations, offline smoke mode, and CI.
Commit: 7462c77 (supersedes v1.0 from Jul 3).

Verify in 10 seconds: curl http://47.93.234.51:8080/api/health · live at https://engram.axiqo.xyz · DevOps scenario: https://engram.axiqo.xyz/?seed=devops

• Judge quickstart: JUDGING.md
• Alibaba Cloud deployment proof: docs/proof-of-deployment.md
• Benchmark + ablations: docs/evaluation.md
• Offline check (no API key): ENGRAM_FAKE_QWEN=1 python3 tests/smoke_offline.py

v1.0 — Qwen Cloud Hackathon Track 1 submission

Frozen submission snapshot for judging.

• Live (Alibaba Cloud ECS Beijing): http://47.93.234.51:8080 · HTTPS mirror: https://engram.axiqo.xyz
• Demo video (narrated, subtitled): https://youtu.be/xZvDm35cXos · Deployment proof: https://youtu.be/DDso1eEqKTo
• Reproducible evaluation: docs/evaluation.md (5/5 scenarios vs no-memory & full-history baselines)
• Alibaba Cloud usage proof: backend/qwen_client.py (also served at http://47.93.234.51:8080/qwen_client.py)
• ECS instance: i-2zefhmpp3htrijv7plwr (cn-beijing-c)