memla 0.1.0

Memla CLI for bounded coding and math runtimes.

Memla CLI

Memla is a bounded runtime that helps smaller local models make better technical decisions inside verifier-backed loops.

This public repo is the CLI-first version of Memla:

• local coding workflow planning and execution
• coding benchmarks for patch execution and compile/backtest loops
• bounded math benchmarks for decision-layer proof
• a small proof packet showing the current strongest results

It is intentionally narrower than the internal research repo.

What is here

• memla.py
  • top-level CLI entry point
• memory_system/
  • the runtime, coding session, patch benchmark, compile benchmark, math benchmark, and pack builders
• cases/
  • small bundled case files for quick benchmark runs
• proof/
  • the current proof packet and key benchmark artifacts
• tests/
  • focused tests for the coding, math, and CLI surfaces

Current claim

The current public proof packet is in:

• proof/current_pack/

Current strongest bounded claim:

• on coding, qwen3.5:9b + Memla beat local qwen2.5:32b raw on execution outcome in the OAuth patch benchmark
• on math, qwen3.5:4b + Memla matched qwen2.5:32b raw on the bounded harder pack
• on ambiguous math decision states, Memla lifted both 4b and 9b to perfect choice accuracy on the tested slice

This is not a claim of universal model parity. It is a claim about bounded runtimes with verifiers.

Quick start

Prerequisites:

• Python 3.11+
• Ollama running locally
• one or more local models already pulled

Install:

py -3 -m pip install .

Smoke-check the CLI:

memla --help

Run a local environment check:

memla doctor --repo-root . --model qwen3.5:9b

Main commands

Build a workflow plan inside a repo:

memla coding plan --prompt "Fix the auth regression" --repo-root .

Run a bounded coding turn with optional verification:

memla coding run --prompt "Repair the failing auth tests" --repo-root . --test-command "pytest -q"

Run the patch execution benchmark:

memla coding benchmark-patch --pack path\\to\\git_history_case_pack.json --raw-model qwen2.5:32b --memla-model qwen3.5:9b

Run the compile-loop benchmark:

memla coding benchmark-compile --cases cases\\coding_eval_cases.jsonl --repo-root . --model qwen3.5:9b

Run the bounded math benchmark:

memla math benchmark --cases cases\\math_linear_c2a_v2_harder.jsonl --teacher-model qwen2.5:32b --student-models qwen3.5:4b qwen3.5:9b --executor-mode stepwise_rerank --teacher-trace-source hybrid

Rebuild the proof packet:

memla pack thesis --coding proof\\current_pack\\frozen\\coding_patch_execution_report.json --math-rerank proof\\current_pack\\frozen\\math_step_rerank_report.json --math-progress proof\\current_pack\\frozen\\math_progress_report.json

Benchmark commands write report bundles under ./memla_reports/ by default.

Publish the current proof pack to the repo root for Vercel:

memla pack publish-site --source proof\\current_pack --out-dir .

Proof artifacts

Main public proof packet:

• proof/current_pack/

Supporting reports:

• proof/reports/coding_oauth_patch/
• proof/reports/coding_fastapi_patch/
• proof/reports/coding_fastapi_compile/
• proof/reports/math_qwen32b_rerank/
• proof/reports/math_qwen32b_stepwise/

Tests

Focused verification:

py -3 -m pytest -q tests\\test_step13_coding_compile_loop.py tests\\test_step14_compile_loop_benchmark.py tests\\test_step15_patch_execution_benchmark.py tests\\test_step16_math_c2a_benchmark.py tests\\test_step17_memla_cli.py

Product direction

Memla is being packaged as:

• a local/private coding runtime for smaller models
• a CLI first, not a chat app first
• a verifier-backed system, not a prompt wrapper

The wedge is simple:

make local 9b/14b/32b coding models more execution-capable than their raw form.

GitHub and Vercel

This repo is set up so GitHub and Vercel can work together cleanly:

• GitHub Actions runs the focused CLI/benchmark test suite on push and pull request
• memla pack publish-site --source proof\\current_pack --out-dir . syncs the current proof pack to the repo root
• the root vercel.json is written as a static deployment config so Vercel can deploy the proof site directly from the repo root

Practical flow:

1. update proof artifacts
2. run memla pack publish-site
3. push to GitHub
4. let Vercel redeploy from the latest commit