cookbook-memory

AI Agent Memory Harness (codename: Cookbook Memory)

Repository: https://github.com/kenhuangus/agent-memory-harness · Live site: https://kenhuangus.github.io/agent-memory-harness/

A project site (GitHub Pages) for a persistent memory harness for long-running coding agents.

The harness gives an agent self-curating memory so a smaller model (Haiku) can close the gap to a frontier model (Opus 4.8) without memory, measured on public memory benchmarks. This repo holds the static website that documents the plan, architecture, benchmarks, implementation contracts, and a results scoreboard.

What's here

Page	File	Contents
Overview	index.html	Hypothesis, the four modules, the four metrics
Plan	plan.html	Problem, technical approach, scope, ownership, Gantt + milestones — with a PDF download
Architecture	architecture.html	The diagram, data flows, the router, indexing
Benchmarks	benchmarks.html	The 5 public benchmarks + metric mapping, with links
Implementation	implementation.html	Schemas, storage interface, router/dreaming contracts, eval protocol
Results	results.html	Live scoreboard rendered from results.json (v0.1 results exist; see results/)
Collaborate	collaborate.html	Branch model, ownership, running the pipeline, PR/merge flow

.
├── index.html  plan.html  architecture.html  benchmarks.html  implementation.html  results.html  collaborate.html
├── project-plan.md          # full plan (problem · approach · scope · ownership · timeline)
├── project-plan.pdf         # downloadable PDF of the plan
├── plan.md · prd.md · architecture.md  # the design contracts (what · why · how)
├── assets/
│   ├── css/style.css        # theme + all components
│   ├── js/main.js           # nav toggle, active link, reveal-on-scroll
│   └── img/architecture.svg # standalone architecture diagram
├── eval/                     # the evaluation harness (Python pkg `memeval`) — see "Run the benchmarks"
├── plugin/                   # the installable Cookbook Memory plugin (cookbook_memory) — see plugin/README.md
├── results/                  # committed benchmark results, bucketed by memory version (v0.1, v0.1-bm25, …)
├── docs/                     # ADRs and design records (docs/adrs/**)
├── tools/                    # helper scripts (dataset probes, etc.)
├── .nojekyll                # serve assets as-is (no Jekyll build)
└── README.md

The site is plain HTML/CSS/JS — no build step, no dependencies. Open index.html locally, or publish it straight to GitHub Pages.

Publish to GitHub Pages

1. Create a repository on GitHub (e.g. agent-memory-harness).

2. From this folder:

   git init
   git add .
   git commit -m "Add AI Agent Memory Harness project site"
   git branch -M main
   git remote add origin https://github.com/<you>/agent-memory-harness.git
   git push -u origin main
   

3. On GitHub: Settings → Pages → Build and deployment → Source: Deploy from a branch, pick main / / (root), and save.

4. The site goes live at https://<you>.github.io/agent-memory-harness/ within a minute or two.

All internal links are relative, so the site works at a project-pages subpath or at a user/organisation root without any changes.

The five benchmarks

• MemoryAgentBench — paper · code · dataset
• LongMemEval — paper · code · site
• SWE-ContextBench — paper · dataset · code
• SWE-Bench-CL — paper · code
• ContextBench (in-task retrieval quality) — paper · dataset · code

Complementary: LoCoMo, SWE-bench Verified.

Dataset schemas & sample records: see benchmark-schema-sampledata.md for each benchmark's source JSON schema, a truncated example record, and how it maps to the harness Task shape.

Run the benchmarks (the eval/ harness)