STRATA v2.2
Self-hosted AI memory server. Your thoughts, your hardware, your data.
Why Strata?
Your AI forgets everything the moment the session ends. Strata fixes that. It gives any MCP-compatible AI a persistent memory that lives on YOUR hardware - not in the cloud, not on someone else's server. Search by meaning, attach real files, and run it all on a Raspberry Pi.
What's new in v2.2
v2.0 taught Strata who's talking. v2.2 teaches it to scale, audit, and remember your past.
- Per-agent API keys - every AI agent gets its own key with granular
read / write / delete / admin / kill permissions. You stop sharing one key across every tool. Manage them from /admin/agents.
- Global MCP kill switch - one toggle that locks every agent out of MCP, REST, and
/api/*. Any agent with write or kill permission can pull the brake; only a human admin can flip it back on. Built for the moment a tool goes rogue.
- 3D Constellation viewer - watch your brain think in real time at
/constellation. Sacred-geometry layout (Flower of Life background, dodecahedron clusters, Fibonacci sphere distribution), per-agent colors, live activity stream, and a kill-switch indicator that desaturates the whole scene when MCP is offline.
- Per-agent identity colors - each agent gets a color from a curated palette, editable from the admin panel. Colors flow through the constellation viewer so you can tell at a glance which agent is doing what.
- Encryption at rest (optional) - SQLCipher AES-256 encryption with key in a root-only file. Dormant by default; flip on with one env var and a setup script when your threat model needs it.
- File-level hardening (optional) - dedicated
strata system user owns the database with mode 600. SSH users can no longer bypass the API by editing the SQLite file directly.
- Demo mode - run a public demo with
STRATA_DEMO_MODE=true and the dashboard accepts blank passwords (the login screen still renders so visitors can see the auth feature exists).
New in v2.2:
- PostgreSQL + pgvector backend — SQLite is still the default for quick demos, but set
STRATA_DB_BACKEND=postgresql for production. PostgreSQL handles concurrent multi-agent writes natively (no more write queue), and pgvector does similarity search IN the database with HNSW indexes. Scales to 1 million+ thoughts.
- CSV audit log — every agent action is logged to daily CSV files in
data/audit/. Full replay tape: who searched what, which thoughts came back, how long it took. This is the transparency layer — trace exactly how an AI built its understanding of your data. View via GET /api/audit.
- Pre-Strata history (negative IDs) — import your memories from BEFORE you installed Strata using
capture_legacy. These get negative IDs (-1, -2, -3...) and live in the same constellation alongside your current thoughts. Your AI memory doesn't start at install day — it reaches back as far as you want. Click "THOUGHTS" in the constellation viewer to toggle between current and historical views.
- 10 thought types — expanded from 8 types with clear descriptions that teach AI agents WHEN to use each one. Added
idea (brainstorms, not yet decided) and observation (patterns noticed, no conclusions drawn). The agent protocol now spells out type selection so every new user's constellation is balanced.
- Updated agent protocol —
strata_status now teaches agents about all 10 types, negative ID history, the audit log, and outcome loop tracking. No CLAUDE.md needed — the protocol IS the onboarding.
- Constellation camera fix — clicking a star flies the camera past it and looks back, keeping the origin star visible in the background as an anchor.
- Demo mode improvements — agent creation and MCP kill switch now work without an admin key in demo mode.
Plus the v1 features you already know: semantic search, file vault, dedup protection, password+seed-phrase dashboard auth, dark theme, runs on a Pi.
I built this because I got tired of my AI tools forgetting everything between sessions. STRATA gives any MCP-compatible AI (Claude Code, Codex CLI, whatever comes next) a persistent brain that lives on YOUR machine. Not in the cloud. Not on someone else's server. Yours.
It runs on a Raspberry Pi 4B. Seriously. The whole thing - semantic search, file storage, embeddings - runs on a $55 computer with 4GB of RAM.
What it does
upstash
anthropics
1270011