Skip to content

SUBMISSION_FORM.md

Latest commit

 

History

History
60 lines (45 loc) · 2.6 KB

SUBMISSION_FORM.md

File metadata and controls

60 lines (45 loc) · 2.6 KB
Raw

AGI Summit 2026 Hackathon — Submission Fields (copy-paste)

Check-in: Jul 18 9:00 AM – Jul 19 5:00 PM. Fields editable until deadline.

1. Project Title

Cogram — Procedural Memory for Coding Agents

2. Project Description

Coding agents (SWE-agent, OpenHands, Cursor, Claude Code) are amnesiacs: every run starts cold and re-derives fixes they already found last month. The usual patch — stuffing history into context — costs tens of thousands of tokens per turn and still degrades on long context.

Cogram is a procedural-memory substrate. It reads an agent's past work (reasoning → command → error → fix → patch) and builds a concept graph with line-level provenance — using ZERO LLM tokens (pure statistics: co-occurrence + a surprise/novelty signal). Recall costs ~200 tokens and points to the exact source line of the past fix. Memory decays by experienced interactions ("ticks"), never deletes — stale fixes fade but re-light on demand.

Target user: anyone building or running coding agents who wants cross-session learning without paying an LLM to memorize.

Validated on 600 real public SWE-agent trajectories (CC-BY 4.0): 0-token learning, 4.5× index compression, 7,000–17,000× smaller recall than re-reading history. On the standard memory benchmark (LongMemEval, ICLR 2025) it matches the best retriever at ~45× fewer tokens. Cloud persistence via InsForge.

3. GitHub Repository

[TODO] create a PUBLIC repo with ONLY the cogram/ code (do NOT link the personal monorepo).

4. Demo Video (< 3 min)

[TODO] record: problem (agents forget) → 0-token learning on SWE trajectories → new error → recall in ~200 tokens with file:line provenance → tick slider (decay-never-delete). Link here.

5. Primary Track

Open / AI-infrastructure (dev tools). Not healthcare or fintech — pick the general/open track.
Sponsor prize target: InsForge (agent-native BaaS, $500).

6. Sponsor Technologies Used

InsForge — used as the cloud-persistence backend for Cogram's concept graph
(Postgres-backed store synced via the InsForge CLI + API; AI Gateway used for benchmark LLM-as-judge).

7. Terms & Conditions

[check the box]

8. Live Project URL (optional)

[optional — web demo currently runs locally at localhost:8090; leave blank or deploy later]

Blockers to clear before final submit

  1. Public GitHub repo — the cogram/ folder only, scrubbed of personal paths. (Cogram code is self-contained; ~15 py files + web/ + README.)
  2. Demo video — ≤3 min screen recording.
  3. (optional) deploy web demo for a Live URL.