These are the artifacts. Posts argue. These prove. Memory, tools, deterministic constraints, evaluation, observability, security, human control.
Memori — flagship
Persistent memory for AI coding agents. Rust core, SQLite, FTS5, vector search. One file. No API dependency.
Problem. Every agent session starts from zero. Context you already paid for is gone the next time you open the repo.
Thesis. Memory is not a notes database. It is three jobs: capture useful state, preserve it cheaply, retrieve the right item at the right moment. If previously learned information cannot be recovered when needed, memory did not happen.
Architecture. Hybrid search with RRF, access decay scoring, cosine deduplication. Brute-force over an index until the collection size forces otherwise.
Implementation. Python bindings, a CLI designed for agents, Claude Code integration, a published package with prebuilt wheels for Linux (manylinux), macOS, and Windows, and a Rust crate.
Results. About 190 tests, benchmarks, and a retrieval loop that actually survives across sessions.
Lessons. Designing the CLI for agents made it better for humans. Hard constraints belong in code, not in the prompt.
Code. github.com/archit15singh/memori · py-memori on PyPI · memori-ai-core on crates.io · docs.rs
Talk. Memory is the Agent, PyCon India 2025.
Research.
Luffy
A comment-triggered PR review control plane. The agent can shell into the workspace, reproduce failures, post a structured review, store a redacted trace, and grow hub memory so the next run on that repo is not starting from zero.
Problem. Most PR bots read a diff, emit soft prose, and vanish. No memory of the last review. No package you can open later to see which tools ran. No honest score when the code is wrong.
Thesis. Agents need state, observability, and a control plane. Not a better prompt.
Architecture. GitHub comment → Actions → Hermes Agent through OpenRouter → tools in the workspace → structured Markdown review → redacted trace → hub memory.
Implementation. You type @luffy review this pr. Model choice is explicit. Memory is per-repo and accumulates.
Results. Live Odoo review on a real bug: REQUEST CHANGES, score 42/100, ten API calls, about four minutes, roughly $0.59. The trace package is in the repo under docs/showcase/.
Lessons. If you cannot audit the loop, you do not have a review agent. You have a comment generator.
Code. github.com/archit15singh/luffy-pr-review-agent
Research. Building Luffy
Also original
GitHub only. No post yet.
Free Model Eval. Six free-tier coding models through OpenCode, eleven variants, ten tasks. Welch t-test, Bonferroni, Cohen’s d. The grader is a different model from the one under test.
Chronicle. Engineering knowledge dies when people leave. Chronicle rebuilds the decision graph from the repo: PRs, architecture, incidents. Document search is the wrong product.
Idea Generator. An OpenCode skill that is a DAG. Ingest YC companies in a constrained market, descend, cut a wedge, score founder-fit. Gates and typed receipts live in idea_factory/. The prompts do not.
Sim Bangalore. 243 BBMP wards, one synthetic agent per resident, A* on road plus Metro, a scorecard for WFH / fare / ORR-closure experiments. The foundation is a port of Sim Francisco. The Bangalore population, calibration, and policy core are the original work.