Hand-curated .memee packs for the stacks your agents
already know. Patterns, anti-patterns, decisions. Signed, sourced,
idempotent. Import in one command. User canon outranks pack defaults
the moment you record your first lesson.
Five packs ship today. Community submissions land via PR to github.com/gizmax/memee. The agent-discipline pack has its own gate — see CONTRIBUTING.
FastAPI, SQLAlchemy 2.0, Pydantic v2, pytest, alembic. Async correctness, observability, deployment. Plus the OWASP-shaped security canon every Python service needs from day one.
Hooks rules, forms, accessibility, animation perf, mobile gotchas. TypeScript strict, Tailwind, TanStack Query, React Router v6 loaders. Most rules generalise to any modern React SPA.
Build MCP servers that don’t get pwned, abandoned, or ignored. Transport, tool design, security (lethal-trifecta read+write split, supply-chain pinning, prompt injection in descriptions), lifecycle, deployment.
RFC 9110 + 7807/9457 + 8725 + 9700 + OAuth 2.1 distilled. Status discipline, problem+json errors, JWT BCP (alg-none, key confusion), OAuth pitfalls (implicit flow, redirect-URI matching), ETag concurrency. The IETF index nobody had written.
Failure modes that fail across every team. Fabricated paths, premature done, scope explosion on small fixes, destructive ops without consent, sycophancy under pushback, the lethal trifecta. Every entry cites primary literature: Anthropic, arXiv 2509.18970, Replit incident, Simon Willison, sycophancy.md. The rules CLAUDE.md keeps re-deriving in every repo, written once.
SwiftUI state, @State / @Observable / @Bindable, gesture pitfalls (DragGesture + offset structural bug), NavigationStack, preview crash modes, Swift Concurrency, UIKit interop.
Language-agnostic Postgres canon. Indexing strategy (partial, GIN, GiST, BRIN), EXPLAIN reading, lock contention, JSONB vs columns, pgbouncer modes, statement_timeout, VACUUM tuning.
A .memee pack is a gzipped tarball: manifest.toml + memories.jsonl + an optional ed25519 signature. Distributable however you ship a file. GitHub releases, S3, an internal artefact store, an email attachment.
The format is documented (docs/pack-format.md) and small. Build your own tooling. The format is the artefact, not the registry.
Every memory in a pack imports with source_type=import (×0.6 confidence multiplier) and a per-pack confidence cap. Your hand-recorded lesson always outranks the pack default once you validate it.
Re-installing the same pack version is a no-op. Re-installing a newer version dedupes against the local canon and folds via the existing quality gate.
A fresh memee setup used to ship empty. Day-one users had no canon to route, no anti-patterns to surface, no decisions to recall. The dogfooding theory took six weeks to populate the DB.
One pack collapses that to one command. memee pack install python-web and you have 61 sourced lessons before your first agent session.
memee pack export --canon-only > team.memee writes everything in your local DB at maturity=validated|canon. Strip identity columns, sign with MEMEE_PACK_KEY, ship.
Internal teams: cross-project canon as a packagable artefact. Agencies: client handover as a file. OSS maintainers: house style as a public download.