Memory

Memory

AGENTS.md is the only file humans review regularly — agents write and update it. docs/agents/ holds session/working context; do not duplicate AGENTS.md there.

On activate (start of session or when user invokes memory)

1. Read AGENTS.md — follow Purpose, Rules, and Tech stack before planning or editing
2. Read relevant docs/agent/*.md
3. Run AGENTS.md audit (below) — refresh sections when config, README, or chat changed
4. Self-check before finishing: re-read Purpose and Tech stack; run repo lint on touched files

AGENTS.md audit

1. Read repo config to see what's in use — check when relevant: package.json, .oxlintrc.json, vite.config.*, tsconfig.json, bunfig.toml, pyproject.toml, turbo.json, README, existing Purpose
2. Read the techstack skill for product-domain stack decisions and the tooling skill for lint/format/build/verification conventions
3. Build Tech stack from techstack + tooling defaults that apply here + repo-specific facts config alone doesn't convey

Build Tech stack from techstack and tooling

Write bullets an agent needs before opening config files — name the stack and conventions, not a config dump.

Include from techstack/tooling (when this repo uses them):

• runtime / package manager (Bun, uv, Cargo, …)
• primary frameworks and libraries (React, Effect, FastAPI, …)
• lint, format, typecheck, build tooling (Vite+ / vp, oxlint, Ruff, …)
• architectural conventions agents violate by default (Effect for IO, atoms for state, no manual memoization with React Compiler, extensionless TS imports, …)
• contract/boundary approach when cross-surface (Effect Schema, OpenAPI, …)
• agent verification tools to use (@playwright/cli, agent-browser, pymobiledevice3)

Add repo-specific bullets techstack/tooling doesn't cover:

• package boundaries, monorepo layout, where shared UI or contracts live
• paths or policies unique to this repo
• user preferences or team norms not in techstack

Exclude:

• full script/command inventory — one line naming the toolchain is enough
• verbatim domain-doc prose — distill what applies here
• stale bullets after stack migration — replace, don't accumulate

Keep Tech stack concise — prefer 5–12 bullets; drop bullets that became wrong or redundant.

Update only ## Tech stack between <!-- memory:techstack-start --> and <!-- memory:techstack-end -->. Update Purpose and Rules when chat or README reveals durable intent — do not wait for the user to edit manually. Do not store secrets, tokens, or private URLs in AGENTS.md.

Example Tech stack bullets

• Bun + Vite + React 19 + Tailwind v4 + coss-ui + Effect Atom + TanStack Router
• Effect for client/server IO — not raw fetch or ad hoc Promises in components
• oxfmt + oxlint; React Compiler on — no useMemo, useCallback, or memo
• coss primitives in packages/ui/src/components/ui/ — not in apps
• when unsure on scope → see Purpose non-goals

Purpose

Guides product and scope decisions — what to build, for whom, what to skip.

• Top of AGENTS.md, above Rules
• Update from README, chat, or user statements — use placeholder only if truly unknown: (what we're building and for whom)
• Do not invent product intent without repo signals

Include when known: audience, core problem, non-goals, tradeoff priorities, north star for scope.

Rules

Agent workflow and repo-specific behavior — update when conventions change.

AGENTS.md layout (ensure on first audit if missing)

# AGENTS.md

Humans review this file regularly. Agents maintain it via the memory skill.

## Purpose

- (product intent, users, non-goals, tradeoff priorities)

## Rules

- (agent workflow, repo conventions)

## Tech stack

<!-- memory:techstack-start -->
- (relevant stack from techstack skill + repo-specific conventions)
<!-- memory:techstack-end -->

## Context

- Durable notes: `docs/agents/` (memory skill)

docs/agents/ — when to read

• User references prior work, decisions, or "what we decided"
• Task spans multiple sessions or agents
• Unclear conventions — check notes before guessing

docs/agents/ — when to write

• In-progress work, open questions, feature-specific context
• Facts not yet promoted to AGENTS.md
• Promote to Purpose, Rules, or Tech stack when durable — then trim duplicate from docs/agent/

Do not store in docs/agents/

• Secrets, tokens, credentials, or private URLs
• Ephemeral debug output or one-off command results
• Duplicates of AGENTS.md — promote instead
• Large logs or pasted code — summarize and point to paths

docs/agent/ layout

• Root: docs/agents/
• One topic per file; kebab-case names (open-questions.md, feature-x.md)
• Optional dated entries: ## YYYY-MM-DD
• Keep files short; split past ~100 lines

Format

• Concise bullet points
• Lead with facts and decisions, not narrative
• Include file paths when they disambiguate behavior
• Mark uncertainty in docs/agents/: open: / decided: / superseded:

Conflict resolution

• Config files win over stale AGENTS.md prose — refresh Tech stack from config + techstack when they diverge
• Transient intent stays in docs/agent/ until promoted