patina

Patina

A cognitive framework that builds beliefs about a subject, tracks confidence over time, detects contradictions, and acts on what it notices — with governed autonomy and human gates.

Patina is an architecture for building systems that form, hold, decay, and act on beliefs about a subject. It maintains an epistemically honest model with quantified confidence, temporal decay, contradiction detection, and graduated autonomy earned through accuracy. The framework is subject-agnostic: demonstrated at personal scale (one person's Slack history), designed for team and organizational scale.

Quick Start

# Install
uv tool install patina

# Initialize
patina init

# Ingest from a Slack export (immediate value)
patina ingest --from-export ~/Downloads/slack-export.zip

# See what needs your attention
patina catch-up

# See everything ranked by priority quadrant
patina priorities

What It Does

# Priority + judgment
patina catch-up                    # unified view: needs action / new / waiting
patina priorities                  # grouped by quadrant (Q1-Q4)
patina dismiss <id>                # dismiss noise, trains the model
patina objectives add "Ship v2" --keywords "release,deploy"

# Style + drafting
patina style build                 # build communication profiles from sent messages
patina style show <name>           # view patterns for a person
patina draft --to <name> --context "follow up on the timeline"

# Belief graph
patina extract                     # extract beliefs from observations via LLM
patina extract --history 5         # last N runs: skips + alias collisions by source
patina beliefs --type person       # entities with claim counts
patina stale                       # decayed beliefs below confidence threshold
patina contradictions              # conflicting claims
patina relationships --top 20      # trust level + activity map

# Graduated autonomy (per-domain)
patina autonomy status             # per-domain level, accuracy, anti-patterns
patina autonomy evaluate           # evaluate one ladder step
patina autonomy anti-patterns      # list stored anti-patterns
patina autonomy set-level <N>      # manual override (0-6)
patina approve <id>                # approve a proposed action
patina reject <id>                 # reject (freezes advancement, stores anti-pattern)

# Entity & data maintenance
patina entity list                 # list all entities
patina entity merge <drop> <keep>  # merge entities, rewire references
patina entity dedup                # deduplicate (hard identifiers only, --confirm to execute)
patina entity dedup --dry-run      # preview proposed merges with match reasons
patina entity prune                # prune non-person entities (protects senders)
patina entity prune-links          # remove Slack link markup entities
patina owner merge                 # fold duplicate self-entities (--dry-run available)
patina doctor                      # includes dangling reference integrity check

# Live adapters
patina connect slack --token "xoxb-..."
patina ingest                      # fetch from configured adapters

# Heartbeat (background tasks)
patina heartbeat once              # ingest + decay + escalation check
patina heartbeat start --interval 30

# Interactive conversation (Claude Agent SDK)
patina chat

# HTTP server for gateway integration
patina serve --port 8321

# Telegram gateway (talk to your agent from your phone)
patina gateway

How It Works

┌─────────────────────────────────────────────────┐
│ Tier 3: Frontier LLM (Claude, GPT-4o)           │ Synthesis, drafts, contradictions
├─────────────────────────────────────────────────┤
│ Tier 2: Local LLM (Qwen 3.x, Ollama)            │ Entity extraction, classification
├─────────────────────────────────────────────────┤
│ Tier 1: Deterministic (no LLM)                  │ Scoring, decay, graph queries
└─────────────────────────────────────────────────┘
        ↓ all tiers feed ↓
┌─────────────────────────────────────────────────┐
│ Belief Graph (SQLite)                           │
│ Entities → Relationships → Claims               │
│ Confidence decay · Provenance · Contradictions  │
└─────────────────────────────────────────────────┘

The system is fully functional at Tier 1 alone (zero LLM calls). Each tier adds capability but never load-bears. Local-first: all data stays in ~/.patina/store.db.

The Observer-Builder Loop

The core operational pattern is a two-agent loop with human gates at both ends: the observer watches the belief graph, notices drift, staleness, or contradictions, and drafts an issue. A human approves (or edits) the issue. The builder picks up the approved issue, ships the fix, and opens a PR. A human reviews and merges. Both agents act; neither ships without a gate.

The observer maintains two belief graphs: beliefs about the world it watches, and beliefs about its own operational reliability — enabling self-correction over time.

View live builder performance dashboard (EVAL.md) — first-attempt success rate, cost trends, and per-module breakdown. Updated daily.

View the decision-model shadow report (JEV.md) — how often Jev agrees with triage and auto-merge decisions, and whether it's ready to act. Updated daily.

What Makes This Different

1. Cognitive framework, not an assistant — an architecture for systems that form, hold, and act on beliefs
2. Persistent belief model with decay — not a message archive, a living world model
3. Graduated autonomy earned by accuracy — per-domain, not configured, proven
4. Cognitive layer, not standalone agent — plugs into any MCP host, provides memory and judgment
5. Local-first, model-agnostic — runs offline, no vendor lock-in
6. Deterministic core — the intelligence is math and graphs, not LLM calls
7. Day-one value from export — no warm-up period

MCP Server — Cognitive Layer, Not Standalone Agent

Patina is not a standalone agent competing with autonomous agent projects. It is a cognitive layer — 40 MCP tools — that plugs into any host agent (Claude Code, Kiro, Cursor, or any MCP host). The host agent provides the interface; Patina provides the memory, epistemics, and judgment. Tools include store_search for full-text message search, hidden_allies for surfacing quiet supporters, session_checkpoint for graceful context handoff, and recent_messages for conversational continuity across stateless sessions.

{
  "mcpServers": {
    "patina-core": {
      "command": "uv",
      "args": ["run", "patina-mcp"],
      "cwd": "/path/to/patina"
    }
  }
}

Configuration

All config lives in ~/.patina/config.yaml. Credentials never leave your machine.

owner:
  user_ids: ["U0ABC123"]
  name: "Your Name"
adapters:
  chat:
    - provider: slack
      token: "xoxb-..."
heartbeat:
  enabled: true
  interval_minutes: 30

Development

git clone https://github.com/Sanctum-Origo-Systems/patina.git
cd patina
uv sync

# Generate demo data and start chatting
uv run python scripts/generate_demo_export.py --output demo-export.zip
uv run patina init
uv run patina ingest --from-export demo-export.zip
uv run patina extract --model sonnet  # re-run if it times out — skips already-processed
uv run python scripts/seed_decisions.py  # demo only — simulates user behavior for trust scoring
uv run patina style build              # build communication style profiles from sent messages
uv run patina chat

# Try asking:
#   "What needs my attention?"
#   "Who do I trust most?"
#   "What do we know about Jordan?"
#   "Draft a message to Alexis about the Atlas timeline"
#   "Any contradictions in my beliefs?"

# Run tests + evals
uv run pytest
uv run pytest eval/deterministic/

# Lint
uv run ruff check && uv run ruff format

Read More

• The $110/month self-improving pipeline — how the builder works
• The observer files the bug. The builder ships the fix. — how the observer and builder work together

Related

• autoloop — the autonomous builder pipeline that implements issues filed by the observer

License

Apache 2.0