What if your AI assistant could notice things — without being asked?
Most AI assistants are reactive. They wait for you to speak, respond to what you said, and move on.
They don't notice that you've been struggling with the same problem for three days.
They don't adjust when you're exhausted at 1am versus focused at 10am.
They answer with equal confidence whether they're certain or guessing.
They respond to your words — not to what you're actually trying to accomplish.
VIGIL is built to close that gap.
VIGIL is an open-source awareness layer — a system of four independent modules that run silently alongside any AI assistant, giving it genuine depth of understanding.
Not memory. Not personality. Not another chatbot wrapper.
Something that lives between conversations. Beneath the words. Always watching. Always learning.
┌─────────────────────────────────────────────────────────────┐
│ VIGIL │
│ │
│ 🔥 EMBER What remains unresolved │
│ 🧭 COMPASS Where you're actually headed │
│ 🌊 TIDE How the user is — right now │
│ 🪞 MIRROR What the AI doesn't know about itself │
│ │
│ [ Your AI Assistant ] │
│ ↑ colored by everything above ↑ │
└─────────────────────────────────────────────────────────────┘
Each module is standalone, zero-dependency, and can be integrated into any AI assistant project — not just FRIDAY.
"Some things don't close. EMBER keeps watching them."
Detects unresolved topics, unfinished tasks, and recurring concerns across conversations. Each ember has a heat level — it glows brighter when related topics surface, fades when resolved, and surfaces proactively when it's been burning too long.
"You asked what. COMPASS knows why."
Tracks the user's actual goal across multiple conversation turns — not the command they issued, but what they're genuinely trying to accomplish. A debugging session is different from a learning session, even when the words look the same.
"It's 1am and you've sent 4 short messages in a row. TIDE knows."
Reads the user's current cognitive and emotional state from behavioral signals — message length, response cadence, vocabulary complexity, time of day — and adjusts the assistant's output depth accordingly. Not long-term style learning. The current moment, right now.
"Knowing what you don't know is the beginning of real intelligence."
Tracks where the AI assistant produces unreliable, inconsistent, or later-corrected answers. Builds a confidence model: which domains, tools, and question types are trustworthy — and which warrant explicit uncertainty signals to the user.
VIGIL is designed to be dropped into any AI assistant pipeline:
# Anywhere in your AI assistant loop:
from vigil import Ember, Compass, Tide, Mirror
ember = Ember()
compass = Compass()
tide = Tide()
mirror = Mirror()
# After each conversation turn:
ember.observe(user_msg, assistant_msg)
compass.observe(user_msg, assistant_msg)
tide.observe(user_msg)
# Before generating a response:
context = ""
context += ember.get_active_context() # "User has 2 unresolved topics"
context += compass.get_goal_directive() # "User is in debugging mode"
context += tide.get_state_directive() # "Keep it short — user seems tired"
confidence = mirror.estimate(query, tool)
if confidence < 0.4:
context += "Acknowledge uncertainty in your response."
# Inject context into your system prompt — done.No framework lock-in. No cloud dependency. No ML training required.
Works with OpenAI, Gemini, Ollama, Groq — or anything else.
VIGIL builds on top of Intelligence-Stones — the cognitive adaptation layer (MindStone, EchoStone, BondStone, IntuitionStone).
Intelligence-Stones → WHO the user is (long-term style, preferences, world model)
VIGIL → WHAT IS HAPPENING right now, beneath the surface
Both systems together: an assistant that knows you and understands the moment.
VIGIL is being actively developed as part of FRIDAY Synapse — a Windows-native AI desktop assistant. Every VIGIL module is validated in production before release.
You get battle-tested code, not theoretical design.
| Module | Design | Implementation | Testing | Release |
|---|---|---|---|---|
| 🔥 EMBER | ✅ | ✅ | ✅ | ✅ |
| 🧭 COMPASS | ✅ | ✅ | ✅ | ✅ |
| 🌊 TIDE | ✅ | ✅ | ✅ | ✅ |
| 🪞 MIRROR | ✅ | ✅ | ✅ | ✅ |
✅ Done · 🔄 In Progress · ⏳ Planned
- Developers building AI assistants or chatbot products
- Anyone integrating LLMs into long-running, multi-session applications
- Researchers exploring persistent context and user-aware AI
- Anyone who's thought "why doesn't my AI just notice that?"
⭐ Star this repository to get notified when modules ship.
Code will be released here — no newsletter, no waitlist, no gatekeeping.
Just good software, when it's ready.
Apache 2.0 — free to use in personal and commercial projects.
"A good assistant doesn't wait to be told everything.
It pays attention."
— VIGIL