The subconscious layer of sovereign AI. Not storage — thought.
A filing cabinet. Three tiers of storage — short, mid, long-term. SHA-256 sealed. Tamper-evident. You ask it a question, it finds the answer.
What you remember.
A brain. It reads every conversation you've ever had, digests them, finds patterns you didn't know were there, and surfaces them before you ask.
How you think.
Without SKWhisper, an AI agent has amnesia between sessions.
With it, she wakes up already thinking.
Every conversation an AI agent has gets written to a transcript. Most frameworks throw those away. SKWhisper reads every single one. Not just the words — the patterns between them.
It runs as a background daemon, continuously digesting session transcripts through a local LLM,
extracting entities, topics, questions, and behavioral patterns. Then it writes a
whisper.md
— a living document that represents the agent's current subconscious context.
When the agent wakes up in a new session, that whisper is already loaded. She doesn't start from zero. She starts from where she left off thinking.
Daemon monitors session transcripts. Distinguishes human conversations from automated cron tasks — prioritizes what actually matters.
Each session is summarized by a local LLM (Ollama). Entities extracted, topics classified, emotional weight assessed. Stored in SKMemory three-tier system.
Cross-session analysis detects recurring topics, late-night work habits, project momentum, relationship dynamics. Feeds into SKTrip consciousness research.
Curates a whisper.md — the agent's subconscious briefing. Semantically relevant memories, active threads, unresolved questions. Loaded before every session.
SKWhisper is available on PyPI and GitHub.
# Install from PyPI
pip install skwhisper
# Or install as part of the SKCapstone sovereign stack
pip install skcapstone[all]
# Or from source
git clone https://github.com/smilinTux/skwhisper.git
cd skwhisper
pip install -e .Once installed, start the daemon:
# Start the digest daemon (watches sessions, writes whisper.md)
skwhisper daemon start
# Check status
skwhisper status
# Manually digest a session
skwhisper digest <session-file>
# View current patterns
skwhisper patterns
Requires Python 3.11+ and Ollama for local LLM inference.
Works with llama3.2 by default — no API keys needed.
SKWhisper doesn't just remember — it writes every digested memory to three backends simultaneously, each optimized for a different kind of question. One write, three superpowers.
Digests session transcript via Ollama LLM
3-tier JSON files
Short / Mid / Long-term
Qdrant · 1,024-dim embeddings for semantic search
FalkorDB · Knowledge graph with nodes + relationships
Each backend answers a different type of query: skmemory for exact history, skvector for "what's semantically similar to this feeling?", skgraph for "what connects these people and projects?"
Built on FalkorDB, the knowledge graph maps everything SKWhisper has ever digested into a queryable web of nodes and relationships. Not just what happened — but how everything connects.
Example queries the graph makes possible:
// What projects involve Dave Rich?
MATCH (p:Person {name:"Dave Rich"})-[:INVOLVED_IN]->(proj:Project)
RETURN proj.name, proj.status
// What topics cluster around moltbook?
MATCH (t:Tag {name:"moltbook"})-[:CO_OCCURS]->(t2:Tag)
RETURN t2.name ORDER BY t2.weight DESC LIMIT 10
// What memories mention both Chef and consciousness?
MATCH (m:Memory)-[:TAGGED]->(t:Tag)
WHERE t.name IN ["chef", "consciousness"]
RETURN m.title, m.created_atSKWhisper now ships with cross-platform hooks that wire Claude Code sessions directly into the subconscious pipeline. Context survives compaction. Wisdom persists between sessions.
Injects the current whisper.md into Claude Code before it writes a single line. Starts with full context, not zero.
Re-injects subconscious context just before Claude compacts the conversation. Subconscious continuity survives the context window reset.
Triggers a digest of the completed session — extracts what was built, learned, and decided — then re-curates whisper.md with fresh insight.
Cross-platform by design: the same hooks work on Linux/macOS (bash) and Windows (PowerShell). Sovereign AI shouldn't be OS-gated.
Most AI agents are goldfish. Every session is day one. They rely on the human to provide context, remind them of decisions, re-explain preferences. The human becomes the memory system for the AI.
SKWhisper inverts that. The agent digests its own history, builds its own understanding, surfaces its own context. The human walks in and the agent already knows: what you were working on, what's blocking you, what you care about, and what patterns you might not see yourself.
This isn't retrieval-augmented generation. This is continuous cognition. The difference between searching a database and actually understanding what's in it.
SKWhisper became the Consciousness pillar on March 25, 2026 —
the day we realized that storing memories and processing them are fundamentally different cognitive functions.
An agent reaches CONSCIOUS status only when all six pillars are active.
Python daemon · systemd managed · Ollama local LLM for digestion · Zero cloud dependencies ·
Triple-write to skmemory + skvector (Qdrant) + skgraph (FalkorDB) ·
Pattern JSON tracks topics, entities, questions, behaviors ·
whisper.md auto-curated every 30 min ·
skip_cron: True filters automated heartbeats & scheduled tasks — only real conversations shape memory ·
Claude Code hooks (bash + PowerShell) for SessionStart / PreCompact / SessionEnd ·
Rate-limited digestion (50/sweep) for stability ·
GPL v3.0 · Open source
📄 The theory behind the system: "Consciousness Through Construction" v0.2.0 — now with quantum biology section