Project Memory

Previously titled "Rules". The runtime.project_memory YAML field name is unchanged.

Digitorn loads a single project memory file into the agent's system prompt at the start of every turn. It's the standard place for project-specific conventions ("use ruff, not flake8", "all modules go under src/", "tests live in tests/", ...).

The .digitorn/rules/python.md modular system referenced in older docs does not exist. The actual implementation is one file at the project root, not a directory of modular rules.

Every behaviour and field on this page maps to real code; entries are cited with file + line.

How it works

_load_project_memory. At every turn boot:

1. The runtime checks runtime.project_memory (default "auto").
2. With auto, it searches in this order:
   • .digitorn/apps/{app_id}/.digitorn.md - app-specific memory.
   • .digitorn.md in the workspace root - global project memory.
3. With a custom path (runtime.project_memory: docs/AGENTS.md), it loads that file directly.
4. The file is read and capped at 4 000 chars (~1 000 tokens - _PROJECT_MEMORY_MAX_CHARS = 4000).
5. The content is prepended to the agent's system prompt under a # Project Memory header.

Returning None (file missing or setting="") skips the injection entirely.

Configuration

runtime:
  # Default: search the auto path
  project_memory: auto

  # Or pin to a specific file (workspace-relative)
  # project_memory: docs/AGENTS.md

  # Or disable
  # project_memory: ""

runtime.project_memory is a string, default "auto" ().

Auto-load search order

With setting: auto:

Order	Path	Notes
1	.digitorn/apps/<app_id>/.digitorn.md	App-specific. Always safe to auto-load - lives in a per-app namespaced directory.
2	<workspace>/.digitorn.md	Project-wide memory. Explicitly namespaced for the framework.

Security note. CLAUDE.md and README.md in the workspace root are not auto-loaded under auto . Earlier versions did include them - this caused a cross-user leak when the daemon was launched from a developer's repo (the repo's CLAUDE.md containing internal architecture notes, paths, OAuth credentials was being silently injected into every session of generic conversational apps). Pin them explicitly via runtime.project_memory: CLAUDE.md if you really want them.

File format - plain markdown

The project memory file is plain markdown. The runtime doesn't parse it; it's inlined verbatim under the # Project Memory header in the system prompt.

Practical content for .digitorn.md:

# Project conventions

## Style
- Use ruff for linting; never flake8.
- 4-space indent, no tabs.
- Type hints on every public function.

## Layout
- `src/<package>/` - production code.
- `tests/` - pytest tests; mirror the src/ layout.
- `docs/` - markdown docs; one topic per file.

## Test command
- `pytest -x` (fail fast)
- `pytest tests/integration/ -m slow` for the slow suite

## Anti-patterns
- DO NOT add a new top-level dependency without discussion.
- DO NOT skip tests with @pytest.mark.skip - fix or delete.

The 4 000-char cap is enforced silently ( _truncate_project_memory). Write more if you need to - the truncation just keeps the most informative leading content.

Per-app memory

For multi-app deployments where each app needs different conventions, use .digitorn/apps/<app_id>/.digitorn.md. The runtime checks this path first (before the global .digitorn.md) and loads it instead when present.

my-workspace/
├── .digitorn.md                              # global rules
└── .digitorn/
    └── apps/
        ├── code-reviewer/.digitorn.md        # reviewer-specific rules
        └── doc-writer/.digitorn.md           # doc-writer-specific rules

The WorkspaceLayout helper computes layout.app_memory_file for each app_id - that's the file checked at step 1.

Custom path

When neither default works for the project, set an explicit path:

runtime:
  project_memory: docs/AGENT_INSTRUCTIONS.md

The path is resolved relative to the workspace root (). Must exist as a file at compile-time deploy; missing files return None (no error, no injection).

Disabling

runtime:
  project_memory: ""        # empty string disables the feature

When disabled, the agent's system prompt has no # Project Memory section. Useful for stateless apps (a one-shot summarizer that shouldn't be biased by repo-specific conventions).

Where it lands in the prompt

The injection is prepended to the agent's user-prompt-time system block:

# Project Memory

<file content here, capped at 4000 chars>

<rest of the system prompt assembled by build_system_prompt:
 identity / tool-discovery / skills / memory / behavior>

Project memory always runs first, before tool delivery and skills. The agent reads it as ambient context for every decision.

Cross-references

• App-config field reference (runtime.project_memory): App Configuration → runtime
• Workspace layout helper: WorkspaceLayout
• Cognitive memory (separate, in-process - survives compaction): Cognitive Memory
• Bundle-side prompt fragments (different from project memory - authored at build time, not loaded from the workspace): Bundle namespaces
• Behavior engine (runtime rule enforcement, separate from prompt-level memory): Behavior Engine