Import

A new memory store is empty, but your knowledge is not: it lives in per-project agent memory files, prose instruction files, READMEs, docs, and years of git history. Import brings that knowledge into zero-memory so briefings are useful from day one instead of after weeks of accumulation.

Two paths cover two kinds of sources. Deterministic import (zero-memory-watcher import) copies existing markdown memory files as-is, with no LLM involved and nothing spent on extraction. LLM bootstrap (zero-memory-watcher bootstrap) takes README/docs pages and the git history, distilled by the extraction pipeline.

Both run on the client machine over the watcher's authenticated transport, and both are idempotent: re-running them only processes what changed.

Deterministic import

File-based agent memory (for example Claude Code's auto-memory directory and the prose sections of CLAUDE.md files) already consists of atomic, curated facts, so it maps almost one-to-one onto memories via the import_memory MCP tool.

Kind comes from each file's frontmatter type: user notes become preferences, feedback becomes conventions, project notes become decisions (when a rationale marker is present) or facts, references stay references. Prose instruction-file sections become conventions.

Scope is routed automatically. User-global files land in your personal scope; project files land in the project scope, resolved from a project hint (the repository root) and auto-created if needed. A project item whose hint resolves badly degrades to your personal scope rather than being dropped, but a project-targeted import with no hint at all is a validation failure, not a degrade. Note the difference from remember, which refuses rather than degrades: the import path predates that contract and keeps its softer behavior for hinted items. See Scopes and isolation.

Provenance is set by the server, never taken from client input. Each imported memory records an import source with the file path and content hash. Because these files are human-curated, imported memories are treated as authoritative, and a later automatic paraphrase of the same fact defers to them. See Provenance.

Idempotency is ledger-based. The content hash of each file is recorded on first import, so a re-run skips everything unchanged, and server-side semantic dedup collapses paraphrases of facts you already stored.

Already-distilled rule files are deliberately not imported: rules are a downstream layer, and importing them would create duplicates flowing against the rules incubator. Source discovery is adapter-based, so support for further tools' memory formats plugs in without server changes.

Flags: --dry-run (preview without writing), --project <hint>, --server <url>, --no-scan, plus --home/--cwd overrides for discovery roots.

Bootstrap from a repository

Bootstrap feeds repository knowledge through the same LLM extraction pipeline that processes session transcripts, with source-aware policies. Docs (root README*.md and docs/**/*.md) are read for standing knowledge — decisions, constraints, conventions — not API reference or boilerplate. History (git log, newest first, default depth 300, batched) is read for the arc of decisions, not individual commits.

Every chunk passes the extraction confidence gate and the per-chunk memory cap. Extracted memories carry provisional bootstrap provenance, the same trust level as watcher-captured memories, so your own writes and confirmations override them naturally.

Usage:

zero-memory-watcher bootstrap [repo-dir] \
  [--dry-run] [--exclude <globs>] [--docs-only | --history-only] \
  [--depth <n>] [--project <hint>] [--server <url>]

• --dry-run is an honest preview: it asks the server's ingest ledger which chunks are genuinely new versus already processed, without claiming anything or spending tokens. If the server is unreachable, the preview degrades to a local listing with unknown verdicts and reports a cost range instead of failing.
• --exclude drops sources by comma-separated globs (matched against the full path and the basename) before any server call, and prints how many were excluded so a trimmed list is visible as trimmed.

A circuit breaker stops the run after three consecutive errors, since every chunk costs an extraction call. help and version work offline, before login or deployment.

After importing

Deterministic import copies file content verbatim — which makes it the most likely path to carry a stray secret into memory. The write-path content guard addresses this class of risk; running a hygiene scan (scan_hygiene) after a large import is recommended so duplicates and conflicts surface early.

Import is one-directional by design: files flow into memory, not back. The reverse direction, memory as reviewable markdown, is Export. The two formats round-trip, so an exported tree can be re-imported idempotently.