How it works

What it generates

.recall/
├── manifest.json       # pointer to the current snapshot + memory files
├── architecture.md      # detected apps/packages/layers, entry points, dependency direction
├── conventions.md       # naming, aliases, linting/formatting, scripts
├── decisions.md         # human-editable decision log
├── features.md          # features backed by structural evidence (controllers, routes, jobs...)
├── glossary.md          # recurring domain terms extracted from identifiers
├── risks.md              # evidence-based risk findings (missing tests, circular deps, etc.)
├── technical-debt.md    # measurable debt findings with evidence and severity
└── snapshots/
    └── latest.json       # the deterministic scan result everything above is derived from

Every generated file wraps Recall's content in markers; anything written outside them is preserved and survives every recall update.

Task-focused ranking

recall context --task "<description>" ranks which files an agent should read first for that task, using deterministic, local, and explainable signals only — no embeddings, no AI provider, no network access:

• filename, path, and exported-symbol matches against the task description
• workspace name matches
• bounded import-graph proximity (direct and reverse neighbors)
• test/production counterpart linking
• term specificity, so a task term that matches nearly every file in the repository is weighted down relative to a rare, specific term

Every ranked file carries a human-readable reason, available in full via --json.

Safety

• Local-first: everything runs on your machine, in your repository.
• No telemetry.
• Does not execute, import, or evaluate your project's source code or scripts.
• Managed writes are restricted to .recall/, .gitignore, and any output path explicitly passed on the command line.