Mneme

Mneme is a private research archive for personal correspondence: preserve the record, find relevant material deterministically, and follow every claim back to its sources. Sending email is out of scope.

The code currently consists of local Python experiments using fictional EML and mboxrd messages. They demonstrate preservation, incremental ingestion, Find and filters, address identities, threads, staleness detection, citations, and export/restore. They are not yet one usable application. The current review identified correctness and recovery defects that must be addressed during consolidation.

Start here

• Project status and test inventory: what exists, evidence, limitations, and current task.
• Roadmap and deep-review checkpoints: M1–M5 outcomes and R1–R5 review points. The next implementation is M1: one local synthetic prototype.
• Latest project review: findings and their planned disposition.
• Agent operating contract: autonomy, review, and consequential approval boundaries. CLAUDE.md loads it for Claude Code.

In Claude, /mneme-work M1 runs a user-authorized milestone through tests and internal review. /mneme-review R1 requests a deep review. The prompts in .claude/commands/ also work as plain task instructions. These are agent commands; a unified Mneme application command is an M1 deliverable and does not exist yet.

To inspect the existing filtered-search experiment's CLI without ingesting data:

PYTHONDONTWRITEBYTECODE=1 /usr/bin/python3 -B experiments/mneme-local-prototype-4/mneme.py --help

On Linux use an already installed Python interpreter appropriate to the tests. Use the commands in project status for verification; no package installation is required for these Python experiments. Experiments accept reviewed synthetic fixtures only, not the user's mail.

Structure

Location	Purpose
experiments/	Existing reference implementations, fixtures, tests, and historical results
docs/PROJECT-STATUS.md and docs/ROADMAP.md	Current state and one implementation sequence
docs/reviews/	Dated project/milestone reviews
docs/decisions/	Durable decisions, with acceptance distinct from proposals
docs/phase-1/	Preservation, threat, corpus, and acceptance requirements
docs/phase-2/, docs/phase-3/	Historical/deferred candidate evaluation and acquisition evidence
tools/handoff/	Disabled handoff experiment; not required for development
.claude/commands/	Reusable work/review prompts

A maintained mneme/ package and top-level tests/ are planned for M1. Historical paths remain intact for reproducibility. Local archives, private exports, credentials, and runtime state do not belong in Git.

Product and preservation commitments

The source archive is the immutable record: source packages, messages, attachments, containers, sidecars, and provider metadata as supplied. Derived data includes indexes, classifications, identity links, annotations, threads, embeddings, and summaries. Machine-derived state must rebuild from source plus versioned rules; user-authored changes require reversible history.

Planned sources include Gmail/Google Workspace, Outlook, Eudora, MBOX, and EML, including sent mail and attachments. Tested LF mboxrd support does not establish support for every provider export or legacy mailbox.

Find remains independent of AI Ask. People and addresses are distinct; inferred relationships must remain distinguishable from evidence. The intended experience is a calm, fast, low-maintenance private research workspace across Mac, iPhone, and iPad. AI, remote access, real-data ingestion, and deployment remain future scoped work.

An Ubuntu home server, Docker, ZFS, and Tailscale are provisional deployment context, not accepted stack choices. Python is the current experiment runtime, not an automatic production decision.

See the charter, product requirements, preservation contract, security principles, and AI privacy policy for the governing requirements.

Forgejo is canonical; GitHub is the Claude work mirror. Check the actual remote URLs in each checkout before publishing.