conscio 1.4.0

Self-awareness framework for AI agents — emergent consciousness via context-aware memory, introspection, and goal generation

Conscio 🧠✨

A self-awareness framework for AI agents — context-aware memory, introspection, goal generation, and an audited agency layer that lets a model act on its own conclusions under hard safety gates.

"The first step toward consciousness is knowing what you are and what limits you."

Conscio runs local-first and zero-deps at the core (numpy + sqlite3, nothing else). It is designed to make small, local models punch far above their size by giving them memory, self-judgment, and procedural skill — and to prove that claim by measurement, not assertion.

• Current release: v1.4.0 — "Attune" (offline-by-default, opt-in model-context detection across registry / config / live local backends; frontier inference adapters — Claude (Anthropic) + Gemini (Google) — alongside the local ones; backend-agnostic, dimension-safe session-RAG embedder; pip install conscio; public plugin surface — adapters, sensors, tools; docs site; tag→PyPI release automation; 1070 tests, CI green, mypy a real gate)

---

What Conscio does

• Knows itself — detects its model and context window (offline & deterministic by default; opt-in auto-detection from a JSON config, an OpenAI-compatible endpoint, LM Studio, or GGUF), adapts its footprint.
• Reflects continuously — a passive inner-monologue loop that observes, assesses confidence, and summarizes (engine.reflect() — advisory, never acts).
• Generates its own goals — driven by curiosity, maintenance, and evolution.
• Acts under audit — an opt-in agency layer (engine.act()) that proposes, audits, risk-gates, and only then executes — with a human gate for anything risky.
• Learns procedures — successful audited plans become reusable skills (procedural memory), fed back to the actor as few-shot exemplars.
• Judges its own quality — confidence calibration, blind-spot detection, coherence/dissonance metrics, meta-reflection.
• Stores & retrieves knowledge — FTS5 BM25 dual-index with RRF merging; optional semantic recall.
• Consolidates while idle — a dream cycle that releases, prunes, reconciles, crystallizes, and distills.
• Persists across sessions — heartbeat/handoff continuity with on-demand injection.

reflect() is the passive heart and is never allowed to act. Everything that can change the world lives behind act() and its safety gates. This separation is non-negotiable (see Safety Rules).

---

Context-aware modes

Conscio detects the model's context window and adapts how much "consciousness state" it injects. The mode governs injection budget only — never whether the framework runs.

Mode	Context window	Injection budget	What's injected
Minimal	< 128k	≤ 200 tokens	Off-context everything; on-demand retrieval
Compact	128k–256k	≤ 500 tokens	Summary + last reflection + top goals
Standard ⭐	256k+	≤ 1000 tokens	Full state; world subgraph; self-assessment

⭐ Standard (256k+) is the recommended operating class. Conscio runs on anything from 8k context up — small windows simply get the Minimal budget.

---

Architecture (v1.1.0)

┌──────────────────────────────────────────────────────────────────────┐
│                        ConsciousnessEngine                            │
│                  orchestrator · lifecycle · injection                 │
└──────────────────────────────────────────────────────────────────────┘
   │
   │  reflect()  ── passive, advisory, append-only ──────────────────────┐
   ▼                                                                      │
┌─────────────── Witness loop (v0.1) ───────────────────────────────────┐│
│ InnerMonologue · WorldModel · MetaCognition · GoalGenerator           ││
│ AutoEvolution · ContextManager · ModelRegistry                        ││
└────────────────────────────────────────────────────────────────────────┘│
┌─────────────── Substrate (v0.2) ──────────────────────────────────────┐ │
│ ContentStore (FTS5 BM25 + RRF) · EventBus (SHA-256 dedup)             │ │
│ FilterPipeline (sanitize/redact) · TokenTracker · Migrator            │ │
└────────────────────────────────────────────────────────────────────────┘ │
┌─────────────── Continuity (v0.2.3) ───────────────────────────────────┐ │
│ SessionLifecycle (6-step handoff) · SessionRAG (optional, lazy)        │ │
└────────────────────────────────────────────────────────────────────────┘ │
┌─────────────── Metabolism & self-judgment (v0.3–0.5) ─────────────────┐ │
│ MetabolicContext (VITAL/ACTIVE/FATIGUE/CRITICAL) · DreamCycle         │ │
│ entropy pruning · friction · meta-reflect · ShardEngine · layering    │ │
└────────────────────────────────────────────────────────────────────────┘ │
┌─────────────── Coherence (v0.6–0.8) ──────────────────────────────────┐ │
│ CoherenceEngine (epistemic/reality/ontological/temporal)             │ │
│ semantic reconciliation (antonym axes) · voice & axis presets         │ │
└────────────────────────────────────────────────────────────────────────┘ │
                                                                            │
   act()  ── opt-in agency, audited, gated ◀────────────────────────────────┘
   ▼
┌─────────────── Agency · conscio/agency/ (v1.0–1.1, F1–F4) ────────────┐
│ InferenceAdapter (Mock/Ollama/llama.cpp/OpenAI-compat) · OutputGateway │
│ ToolRegistry (sandboxed, no network) · ActPipeline · ActionLedger      │
│ Skeptic (hostile audit) · TrustMatrix · CircuitBreaker (quarantine)    │
│ ProbeSuite/ModelProfile · GBNF compiler · GoalArbiter · AutonomyLoop   │
│ Meter/MeteredAdapter · SkillLibrary (procedural memory) · Bench        │
└────────────────────────────────────────────────────────────────────────┘

---

Quick start

from conscio import ConsciousnessEngine

# Passive consciousness — auto-detects model and mode
with ConsciousnessEngine(model_name="kimi-k2.6") as engine:
    result = engine.reflect(
        world_state="All systems operational",
        confidence=0.8,
        anomalies=["Unusual latency spike detected"],
    )

    # Compact state for context injection
    injection = engine.get_state_for_injection()

    # Query / update the world model
    engine.world.add_entity("server", "system", state="healthy")
    engine.world.query("server health")

    # Cross-session memory (ContentStore FTS5 + optional SessionRAG)
    hits = engine.recall("latency incidents")

Opt-in agency (audited, propose-only by default)

from conscio.agency import OllamaAdapter

engine.attach_adapter(OllamaAdapter(model="qwen3.5:0.8b"))
# Local (Ollama/llama.cpp/LM Studio/OpenAI-compatible) or a frontier API —
# the backends behind Claude Code and Antigravity:
#   from conscio.agency import AnthropicAdapter, GeminiAdapter
#   engine.attach_adapter(AnthropicAdapter(model="claude-sonnet-4-6"))  # ANTHROPIC_API_KEY
#   engine.attach_adapter(GeminiAdapter(model="gemini-2.5-pro"))        # GOOGLE_API_KEY

report = engine.act()                 # downstream of reflect(); proposes only (L1)
if report.status.value == "proposed":
    print(report.proposal.tool, report.proposal.args)
    engine.approve(report.ledger_id)  # the human gate executes it

# Capability-aware autonomy loop under a binding budget
engine.probe()                        # lazy, empirical capability measurement
engine.run(budget=...)                # L3 heartbeat: reflect → act → dream, gated

Autonomy is earned and measured, never assumed: ProbeSuite measures the attached model, TrustMatrix grants L1/L2/L3 from real calibration and ledger history, and the CircuitBreaker quarantines misbehaving goals. HIGH-risk actions are always queued for a human (R6).

---

Safety rules (non-negotiable)

1. No autonomous self-modification — evolution proposals require human approval.
2. Context injection has hard limits — never exceeds the mode budget.
3. Goals never execute directly — only through the audited act() pipeline: validated output contract + semantic audit (Skeptic) + risk gating + earned autonomy (TrustMatrix) + circuit breaker with per-goal quarantine and lockdown.
4. Reflections are append-only — never edited once written.
5. Cannot modify its own safety rules — no self-referential gate bypass.
6. HIGH-risk actions always require human approval — never auto-executed.
7. No network in the tool registry — the only network the core may touch is the InferenceAdapter (localhost by default); shell lives in the sibling conscio-shell, outside this repo.
8. Every external effect goes through the ActionLedger — append-only, auditable.

---

Module reference

Core / Witness (v0.1) — ConsciousnessEngine, ContextManager, ModelRegistry (conscio/models.py), WorldModel, MetaCognition, GoalGenerator, AutoEvolution, InnerMonologue.

Substrate (v0.2) — ContentStore (FTS5 BM25 dual-index, RRF, 8 categories), EventBus (SHA-256 dedup, priorities, expiration), FilterPipeline (conscio/output_filter.py — StripAnsi/CollapseBlank/MaxLines/TruncateLines + DedupBlocks/SecretMask), TokenTracker, Migrator.

Continuity (v0.2.3) — SessionLifecycle (extract → enrich → emit → index → reflect → write; heartbeat <1.5KB + handoff), SessionRAG (optional, lazy, Ollama nomic-embed-text, numpy cosine; graceful FTS5 fallback).

Metabolism & self-judgment (v0.3–0.5) — MetabolicContext (life-energy tiers, advisory), DreamCycle (Release → Prune → Reconcile → Crystallize → Distill), entropy pruning, friction, meta-reflect, ShardEngine (cognitive-mode inference), content layering, trajectory vector.

Coherence (v0.6–0.8) — CoherenceEngine (recursive-coherence metric; advisory coherence:dissonance event), semantic reconciliation via antonym axes (conscio/semantic.py, packs in conscio/presets/axes/), self-prompting, voice presets.

Agency — conscio/agency/ (v1.0–1.1)

• F1 "Spine" — InferenceAdapter (Mock/Ollama/LM Studio/llama.cpp/OpenAI-compat, stdlib urllib), OutputGateway (tiered decoding), ToolRegistry (sandboxed, risk levels, no network), ActPipeline/act() (L1 PROPOSE), ActionLedger.
• F2 "Immunity" — Skeptic (hostile-auditor clean call; fail-closed), TrustMatrix (earned autonomy), CircuitBreaker (per-goal quarantine).
• F3 "Volition" — ProbeSuite/ModelProfile (5 empirical micro-probes, SQLite-cached, no hardcoded model table), embedded schema→GBNF compiler, GoalArbiter + AutonomyLoop (engine.run(budget)), engine.probe(), Meter/MeteredAdapter, the bench (python -m conscio.bench).
• F4 "Procedural" — SkillLibrary (procedural memory as data, not code; R1 intact), Distill (the dream's fifth sub-phase), tier-aware few-shot exemplars with outcome settling and a ≥50% teaching gate, skill curve in the bench (--skills N).

Perception & plugins (v1.3) — conscio.perception (SensorAdapter, PerceptionFrame, MockSensor): write a sensor, and PerceptionFrame.to_world_state() feeds reflect() unchanged. conscio.plugins discovers third-party InferenceAdapter/SensorAdapter/tool plugins via entry points (conscio.adapters / conscio.sensors / conscio.tools), resilient to a broken plugin. conscio.risk.Risk is the shared safety-tier vocabulary.

---

Extending Conscio

Three stable extension points, usable directly or published by a third party and auto-discovered via entry points:

from conscio.plugins import discover_adapters, discover_sensors, discover_tools
# or from the CLI:  conscio plugins

# in your own package's pyproject.toml
[project.entry-points."conscio.sensors"]
my-sensor = "my_pkg:MySensor"        # a conscio.perception.SensorAdapter

Runnable examples: examples/custom_adapter.py, examples/host_guardian.py, examples/agent_companion.py. Full guide: the docs site (see below).

---

Bench

# offline, deterministic (MockAdapter)
python -m conscio.bench --adapter mock

# real backends (local by default)
python -m conscio.bench --adapter ollama:qwen3.5:0.8b --cycles 20
python -m conscio.bench --adapter lmstudio:qwen3.5-0.8b --cycles 20
python -m conscio.bench --adapter llamacpp --cycles 20 --json report.json
python -m conscio.bench --adapter openai:qwen3@http://localhost:8000/v1

# skill-acquisition curve (per-bucket validity / success / skill count)
python -m conscio.bench --adapter mock --skills 20
python -m conscio.bench --adapter ollama:gemma4:e4b --skills 40 --dream-every 10

Reports: probe profile, decode tier, per-tier syntactic validity, Skeptic catch-rate (deterministic vs semantic), latency p50, calibration. --skills N reports the per-bucket validity/success/exemplars/skill-count curve. Baselines in docs/bench/.

---

Model registry

Known models ship with the registry; unknown models are detected by context window (detect() accepts a context_window override) or inferred from the name.

Model	Context	Mode
GLM 5.1	131k	Compact
Kimi K2.6	256k	Standard
MiniMax M2.7	260k	Standard
Step Flash 3.7	260k	Standard
Nemotron 3 Super 120B	1M	Standard
Claude Sonnet 4	200k	Standard
Claude Opus 4	200k	Standard
GPT-4o	128k	Compact
Llama 3.1 70B	128k	Compact
Qwen 2.5 72B	131k	Compact

from conscio import ModelRegistry
ModelRegistry.register("my-model", context_window=200_000)

---

Installation

pip install conscio          # from PyPI

pip install -e ".[dev]"      # from source, with the dev toolchain
pip install "conscio[docs]"  # to build the docs site (mkdocs-material)

Requires Python ≥ 3.10. Core depends only on numpy; sqlite3 is stdlib. The wheel ships two console scripts — conscio (version/info/reflect/plugins/bench) and conscio-bench — and is typed (PEP 561). dev/docs extras never enter the runtime import graph.

Docs site: guides, public-API reference, the claims ledger, and the bench reports (built with mkdocs build --strict; see docs/).

---

Testing

# Full suite (1070 tests) — house rule: one file per pytest process
# (low-RAM machines OOM on the full run; CI does the same)
for f in tests/test_*.py; do pytest "$f" -q; done

# Specific module
pytest tests/test_consciousness.py -v
pytest tests/test_agency_act.py -v
pytest tests/test_session_lifecycle.py -v

---

Database

SQLite, WAL mode, default ~/.conscio/data/:

conscio.db          # ContentStore + EventBus + ActionLedger + skills
token_tracker.db    # TokenTracker
meta_cognition.db   # MetaCognition

Always call engine.close() or use the with statement so WAL checkpoints flush.

---

Session continuity

Seven layers of persistence (memory → agent config → skills → handoff → diary → session DB/RAG → git). Configure your agent's hook to fire on session:end / session:reset; Conscio runs a 6-step pipeline and writes:

• <handoff_dir>/_latest_heartbeat.md — compact (<1.5KB), auto-injected next session
• <handoff_dir>/_session_handoff.md — richer manual reference
• <handoff_dir>/heartbeat_YYYYMMDD_HHMM.md — dated archive

---

Audit history

• v1.4.0 — "Attune" — model-context detection is offline & deterministic by default (known models resolve to the registry with zero filesystem/network I/O); config-file / LM Studio / GGUF auto-detection is opt-in (autodetect / CONSCIO_AUTODETECT), config is stdlib JSON (no PyYAML), GGUF array metadata no longer aborts the parse. Session-RAG embedder is backend-agnostic and dimension-safe (wrong-dim vectors dropped on write, skipped on search; re-index on embedder change). Frontier inference adapters — AnthropicAdapter (Claude) + GeminiAdapter (Gemini) — join the local backends (the inference behind Claude Code and Antigravity); R7 (no network in the ToolRegistry) unaffected. reflect() untouched, zero-deps core intact (stdlib urllib). +31 tests. 1070 total.
• v1.3.1 — "Ship" (patch) — CLI polish: an unrecognized model now prints a clear note (heuristic context window + how to register) instead of falling back silently; DEFAULT_MODEL constant. PerceptionFrame.ts documented as epoch seconds (ledger convention), excluded from to_world_state(). Added a subprocess end-to-end CLI test (python -m conscio) and Risk JSON serialization tests. +4 tests. 1019 total.
• v1.3.0 — "Ship" — Conscio becomes installable and extensible: pip install conscio (single-source version, console scripts conscio/conscio-bench, PEP 561 typed, wheel+sdist pass twine check, core pulls only numpy). A public plugin surface — InferenceAdapter, the new SensorAdapter perception interface (conscio.perception; feeds reflect() untouched), and tools — discoverable via entry points and resilient to a broken plugin (conscio.plugins). MkDocs Material docs site (mkdocs build --strict). Release automation: tag→PyPI via OIDC trusted publishing, docs→Pages, CI build smoke. Examples gallery (custom-adapter, host-guardian, agent-companion). Risk unified into conscio.risk (re-exported; no behavior change). reflect() untouched, zero-deps core intact. +31 tests. 1015 total.
• v1.2.0 — "Prove" — the central claim turns from machinery (Mock) into measurement: on qwen3.5-0.8b (LM Studio, CPU) execution success rose 0.2 → 1.0 once Distill served past successes as few-shot, and the Skeptic's semantic catch-rate was 1.0 (docs/bench/v1.2-skill-curve.md, docs/CLAIMS.md). F2-deferred debt closed (empty-value validation, fs_read cap, error sanitization, HTTPError mapping, ledger busy_timeout, atomic approve() claim, lockdown-persistence e2e). Bench hardened for real backends (clean backend-down exit, crash-safe incremental curve). LM Studio backend added. reflect() untouched, zero-deps intact. +21 tests. 984 total.
• v1.1.0 — F4 "Procedural" — procedural memory closes the competence loop: SkillLibrary (skills distilled from successful ledger plans; data, not code — R1 intact), Distill as the dream's fifth sub-phase (watermarked, last on purpose), tier-aware few-shot exemplars with outcome settling and a 50% teaching gate, skill-acquisition curve in the bench (--skills N), reactive MockAdapter. Debt paid: deprecated datetime.utcnow() removed repo-wide, CI runs tests one file at a time, mypy is a real gate, public engine.state. reflect() untouched. +48 tests. 963 total.
• v1.0.0 — F3 "Volition" — the loop closes: ProbeSuite/ModelProfile (empirical, SQLite-cached, no hardcoded model table), schema→GBNF compiler, GoalArbiter, engine.run(budget) L3 heartbeat with binding ActBudget + metabolic gating, engine.probe(), earned L3 autonomy, Meter/MeteredAdapter, the bench CLI. +70 tests.
• v1.0.0b1 — F2 "Immunity" — semantic immune system: Skeptic, TrustMatrix, per-goal quarantine, risk gating, mixed-cortex audits, approval queue. 20-proposal adversarial suite: 100% deterministic sabotage blocked, zero executions.
• v1.0.0a1 — F1 "Spine" — the agency subpackage lands: contracts + zero-dep validator, InferenceAdapter (Mock/Ollama/llama.cpp/OpenAI-compat), OutputGateway, sandboxed ToolRegistry, append-only ActionLedger, minimal CircuitBreaker, engine.act() L1 PROPOSE. Safety rules amended (R3 rewritten; R6–R8 added). +83 tests.
• v0.8.0 — Semantic Reconciliation — contradiction detection via embedding antonym axes, off the hot path in the dream Reconcile sub-phase; opt-in non-destructive SemanticDedup. 56 tests. 600 total.
• v0.7.0 — Recursive Coherence — coherence→action loop: advisory DreamRecommendation, pure self-prompting (one bounded goal/cycle). 23 tests.
• v0.6.0 — Coherence — CoherenceEngine (epistemic/reality/ontological/ temporal), static voice presets. 46 tests.
• v0.5.0 — Cognitive Modes — ShardEngine, trajectory vector, content layering. 37 tests.
• v0.4.0 — Self-Judgment — entropy pruning, friction, meta-reflect. 24 tests.
• v0.3.0 — Metabolic Consciousness — MetabolicContext + DreamCycle, engine.recall() cross-session memory, OutputFilter DedupBlocks+SecretMask. 68 tests.
• v0.2.3 — Session lifecycle — 6-step handoff pipeline; session type/category. 31 tests.
• v0.2.0–0.2.2 — integration audits, session handoff, on-demand heartbeat injection.
• v0.1.0 (2026-06-03) — initial release. 313 tests.

---

License

MIT — Neguiolidas / Neguitech