yaams 0.6.0

Yet Another Agent Memory System

YAAMS

Yet Another Agent Memory System. A local-first, high-recall memory store for your personal digital exhaust - messages, email, calendar, GitHub activity, notes - searchable in seconds and synthesizable into grounded, cited answers by the LLM of your choice.

YAAMS is Tier 1 of a two-tier memory architecture: the firehose. Tier 2 is the cognitive-ledger: curated atomic notes you keep forever. YAAMS ingests everything, lets you query across the lot, and promotes the gems upstream when you're ready.

Each tier is split between a public engine and a private store:

Tier	Engine (public)	Store (private, your data)
Tier 1	YAAMS - this repo	a single SQLite file at db_path
Tier 2	cognitive-ledger - sibling repo	a markdown tree of curated atomic notes

Engines ship code, no data. Stores live outside the repo, are gitignored by default, and never get pushed.

What

                                 query
                                   |
                                   v
   ingest -> normalize -> embed -> retrieve -> fuse -> answer (cited)
      |          |          |                            |
  iMessage    SQLite    BGE-M3                      Tier 2 boost
  Signal     + FTS5    + spaCy                   (cognitive-ledger)
  Email      + vec
  Teams
  Calendar
  GitHub
  Obsidian

• Local-only by default. Embeddings, NER, and LLM synthesis run on your machine. Nothing leaves the host unless you point an adapter at a hosted backend.
• Append-only. Raw items are immutable. Re-ingest is idempotent.
• Cited answers. Every synthesized claim points back to the source items.
• Pluggable LLM. claude, codex, ollama, any subprocess, or dummy.
• Single SQLite file. No daemons, no servers, no cloud bill.

Why

Language models forget everything between sessions, and the answer is not "shove more raw chat logs into the context window." The answer is a memory system you can grep, query, and audit, that lives on your machine and that you fully control.

YAAMS gives you that for the high-volume side of your life: every iMessage, every email, every calendar event, every GitHub issue you have touched. It normalizes them into a common schema, embeds them, and lets you ask questions like "what did we decide about the deploy in May?" and get back a grounded answer with citations - in seconds, offline.

Quickstart

Requires Python 3.11+ and macOS (Linux works for everything except the iMessage adapter).

Homebrew (recommended):

brew install damsleth/tap/yaams

PyPI via pipx:

pipx install yaams

Bleeding-edge from main:

brew install --HEAD damsleth/tap/yaams

Then bootstrap a config and run the dry-run ingest:

mkdir -p ~/.config/yaams
cp "$(brew --prefix)/share/yaams/config.yaml.example" ~/.config/yaams/config.yaml
$EDITOR ~/.config/yaams/config.yaml
yaams init-db
yaams ingest --dry-run

If you cloned the repo instead of installing the package, run the all-in-one bootstrap script (scripts/install_phase_a.sh) - it creates .venv, runs the config wizard, downloads the spaCy NER model, and runs init-db + ingest --dry-run in sequence.

Verify the setup:

yaams --version          # 0.1.1
yaams stats              # zero items - that is expected before first ingest
yaams ingest --dry-run   # see what each adapter would pick up

Then do a real refresh:

yaams refresh

The first run downloads the embedding model (BAAI/bge-m3, ~2GB). YAAMS will prompt before downloading; after the cache is populated, subsequent runs are fully offline. refresh runs ingest, then safe entity maintenance and association rebuilds.

Ask a question:

yaams query "what did we decide about the deploy in May"
yaams query --answer "open items from the kickoff meeting"

CLI

yaams <command> [options]

Bare yaams lists commands. Global flags: --version, --help. Per-command help: yaams <command> --help.

For the full feature walkthrough — every command, the entity-curation workflow, query flags, and best practices — see the User Guide.

command	what it does
init-db	create the SQLite schema (idempotent)
ingest	run ingest for all enabled sources (or --source <name>)
refresh	run ingest plus safe entity maintenance and association rebuild
curate	run the human entity curation workflow
stats	print item counts per source, plus last ingest run timing
query	full-text + vector search; --answer for synthesized response
feedback	log relevance signals against a prior query result
signals	inspect recent query history
consolidate	group conversational items into sessions
promote	review and accept atomic-note candidates into the Tier 2 ledger
reset-db	drop and recreate the database (asks first)
version	print version (--json for machine-readable)

Ingest sources

source	config key	what it ingests
imessage	ingest.imessage	iMessage conversations from local chat.db
signal	ingest.signal	Signal Desktop messages (1:1 + groups, attachment metadata)
email	ingest.email	.emlx (Apple Mail) or .mbox
notes	ingest.notes	Obsidian vault markdown
tier2_ledger	ingest.tier2_ledger	curated atomic notes from cognitive-ledger
github	ingest.github	GitHub issues and PRs across your repos
calendar / calendar_<profile>	ingest.calendar	Outlook calendar via owa-cal
teams / teams_<profile>	ingest.teams	Microsoft Teams via Graph API

Run one source: yaams ingest --source imessage.

Query

yaams query "budget discussion"
yaams query --top-k 20 "deploy"
yaams query --no-vector "fast FTS-only path"           # skips embedder load
yaams query --source imessage "holiday plans"
yaams query --since 2025-06-01 --until 2025-09-01 "summer"
yaams query --no-consolidations "raw items only"
yaams query --format json "search term"
yaams query --answer "what are the open items from the kickoff"

Promote

Surface candidate atomic notes from recent items, review them interactively, and write accepted ones into your ledger inbox:

yaams promote generate            # scan last 30 days
yaams promote generate --days 60
yaams promote review              # interactive: a/e/r/s/q

Nothing is promoted without your explicit acceptance. Accepted notes land in promote.inbox_path (default ~/yaams/ledger-inbox/).

Configure

config.yaml lives in the repo root (or ~/.config/yaams/config.yaml, or wherever $YAAMS_CONFIG points). It is gitignored: it carries your personal entity dictionary, source paths, and addresses. Edit config.yaml.example instead when contributing structural changes.

Rerun the wizard at any time:

python scripts/configure_phase_a.py --config config.yaml

Key blocks:

db_path: ~/yaams/data.db

ingest:
  since: '2025-01-01T00:00:00Z'
  imessage:
    chat_db_path: ~/Library/Messages/chat.db
  email:
    sources:
      - type: emlx
        path: ~/Library/Mail/V10

embed:
  model: BAAI/bge-m3
  device: mps                # or cpu

synth:
  backend: claude            # claude | codex | ollama | subprocess | dummy
  model: claude-sonnet-4-6

On Apple Silicon, use the Homebrew arm64 Python explicitly - PyTorch 2.4+ has no x86_64 macOS wheels:

/opt/homebrew/bin/python3.12 -m venv .venv

Scheduling

YAAMS is meant to run unattended. See docs/scheduling.md for a launchd agent that runs a single nightly yaams ingest across all enabled sources, plus the macOS Full Disk Access setup required for the imessage adapter to work under launchd.

Privacy and security

YAAMS reads and stores sensitive personal data. The defaults keep everything local, but you are responsible for protecting the database file at rest - enable full-disk encryption (FileVault / LUKS / BitLocker) on the host machine.

See SECURITY.md for the full threat model, data classification, and vulnerability disclosure flow. See docs/privacy-security.md for the operational detail on what is written, what is not, and how to scrub.

Documentation

• User Guide - the end-to-end manual: every feature, the entity-curation workflow, query power-flags, best practices, and troubleshooting.
• Implementation status - architecture and what's shipped.
• Scheduling - unattended nightly ingest via launchd.
• Privacy and security - what's written, what isn't, and how to scrub.

Contributing

See CONTRIBUTING.md for development setup, the test suite, schema-migration rules, and commit conventions.

License

MIT. Copyright 2026 Carl Joakim Damsleth.