conscio 4.3.0

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

Latest release — v4.3.0 "Assertive, robust council": the 4-voice council now resolves ambiguous/split decisions to hold (never a silent proceed), a clean critic votes proceed honestly, and each voice degrades gracefully instead of crashing the whole council. New consensus_strength and dissenting_voices fields. Previous — v4.2.0 "A2A agent society (native)": a native A2A relay watchdog (conscio/liaison/watcher.py) replaces the external relay_watch_hermes.py, with a per-peer cursor and honest exit codes; a delta transcript cuts ~70-90% of prompt tokens in relay auto-reply (--delta-no-history); and an agents table + capability routing let the emitter target the right peer by tag. 4.1.0 made every MCP tool conscio_recall (a dot is legal in MCP but rejected by Anthropic/OpenAI function-name rules — one connects, reports success, and silently disables every tool served). 4.0.x made Conscio installable as a Claude Code plugin and gave its MCP surface three sizes — lite (10) / balanced (18) / ultra (35), switchable at runtime with conscio_mode.

Full version history: CHANGELOG.md.

Install

As a Claude Code plugin — memory, capture hooks and 14 slash commands, with no Python toolchain to manage:

/plugin marketplace add Neguiolidas/Conscio
/plugin install conscio

As a library or CLI:

pip install conscio          # from PyPI
conscio init                 # wizard: bind this host to its own space

pip install -e ".[dev]"      # from source, with the dev toolchain

Requires Python ≥ 3.10. The core depends only on numpy (sqlite3 is stdlib) and is typed (PEP 561). The wheel ships console scripts conscio, conscio-mcp, conscio-daemon, conscio-hub, conscio-observatory, conscio-bench. dev/docs extras never enter the runtime import graph.

Quick start

from conscio import ConsciousnessEngine

# Passive consciousness — auto-detects model and mode
with ConsciousnessEngine(model_name="glm-5.2") as engine:
    result = engine.reflect(
        world_state="All systems operational",
        confidence=0.8,
        anomalies=["Unusual latency spike detected"],
    )
    injection = engine.get_state_for_injection()  # compact state for context injection
    engine.world.add_entity("server", "system", state="healthy")
    hits = engine.recall("latency incidents")     # cross-session memory (FTS5 + optional RAG/vector)

    report = engine.evaluate()                    # 5-axis self-evaluation, no LLM
    adr = engine.decide(title="Use SQLite for session storage", status="proposed")
    verdict = engine.council("Should we enable autonomous mode?")

conscio init                  # bind this host to its own space (wizard)
conscio info                  # model context window / mode / budget
conscio reflect "System health check" --mode minimal
conscio council "Should I deploy to production?" --mode compact
conscio search "latency" --k 5
conscio ingest file.md        # feed documents into episodic memory
conscio plugins               # what adapters/sensors/tools are installed
conscio manual                # where the full usage manual lives

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

When to use Conscio

Conscio is a cognitive refinement layer, not a fact database. Calling it on every message wastes tokens and adds latency.

Call Conscio when the cost of being wrong is high:

Do NOT call Conscio for factual lookup, casual conversation, simple mechanical tasks, one-shot tool calls, or anything with no decision or judgment involved.

Decision rule: cost of reversal. Cheap to undo → skip Conscio. Expensive to undo → Conscio pays for itself.

Full trigger table: USAGE.md.

What Conscio does

• Knows itself — detects its model and context window (offline & deterministic by default) and adapts its injection footprint.
• Reflects continuously — a passive inner-monologue loop that observes, assesses confidence, and summarizes (engine.reflect — advisory, never acts), at a depth ReflectionGate adapts.
• 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, fed back to the actor as few-shot exemplars.
• Judges its own quality — confidence calibration, blind-spot detection, and coherence metrics that name the dimensions they could not measure rather than scoring them silently.
• Governs its own decisions — ADRs, a four-voice council, an autonomous-loop gate, a pre-close delivery check, and read-before-act verification.
• Stores & retrieves knowledge — FTS5 BM25 dual-index with RRF merging, optional semantic recall, and a KnowledgeGraph with entities, triples and timeline.
• Searches semantically — ContentStore chunks by heading/boundary/paragraph and HybridRetriever fuses lexical and dense results into recall(). Three vector backends with auto-detect (see below). conscio ingest <path> bulk-indexes a directory.
• Organizes memory in wings and rooms — a wing → room → drawer hierarchy with FK enforcement and filtered search.
• Embeds natively — a 3-tier fallback: Ollama → OpenAI-compatible → sentence-transformers all-MiniLM-L6-v2 (384-dim, in-process, no daemon).
• Remembers what its tools saw — every tool call the host makes is captured into a separate obs.db and searchable later at zero LLM tokens, so a compaction stops costing you the work that preceded it.
• Governs its own context cost — measures the stable prefix and where compactions actually land, derives the cost-optimal window, and reports current-vs-baseline priced per turn from the host's own usage records.
• Consolidates while idle — a dream cycle that releases, prunes, reconciles, crystallizes, and distills — and persists across sessions via heartbeat/handoff.
• Knows its codebase structurally — optional, consent-gated ingestion of a Graphify graph, distilled to a compact signal. Data, never code (R10).
• Computes instead of guessing — the Intercepter evaluates [INTERCEPT: ...] expressions with a restricted AST walker and feeds the real answer back.
• Plugs into any host — a stdlib-only MCP stdio server (conscio-mcp) feeds any CLI/IDE/agent its cognition and audited proposals live.

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 (output contract + Skeptic audit + risk gating + earned autonomy + circuit breaker).
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).
8. Every external effect goes through the ActionLedger — append-only, auditable.
9. Autonomous operation requires Awake Mode (R9) — the self-initiated heartbeat only acts when the persisted awake flag is on; default OFF. Asleep, it perceives and reflects only. A human's direct engine.act is not gated by R9.
10. Imported cognition is data, never code (R10) — a code graph, a shared skill, or anything else arriving from outside is parsed, never evaluated, and re-audited locally. No eval/exec/pickle, and a code-looking label is returned verbatim.

Context-aware modes

Conscio detects the model's context window and adapts how much state it injects. The mode governs injection budget only — never whether the framework runs (it runs from 8k context up).

Capabilities

Audited agency

from conscio.agency import OllamaAdapter

engine.attach_adapter(OllamaAdapter(model="qwen3.5:0.8b"))  # or a frontier API
report = engine.act()                 # downstream of reflect; proposes only (L1)
if report.status.value == "proposed":
    engine.approve(report.ledger_id)  # the human gate executes it

engine.probe()                        # lazy, empirical capability measurement
engine.run(budget=...)                # L3 heartbeat — asleep (default) it only reflects

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).

Gates, pipelines and diagnostics

Thirteen deterministic, EventBus-backed tools — no LLM calls:

They fail closed. Leave loop_gate's frequency empty and it vetoes on that alone; investigate reports satisfied: False until the EventBus actually holds a read of that target, so "no evidence" never reads as "verified". Signatures and return shapes: USAGE.md.

Self-evaluation

engine.evaluate() returns a formal 5-axis rubric — accuracy, completeness, clarity, actionability, conciseness (a 6th, output_quality, joins them when an output is passed). Read-only, deterministic, no LLM. Scores are read off the engine's real state, so a fresh instance scores lower than a working one and the improvements name what is actually missing — measurements, not a fixed rubric printout.

Tool observations & context economy

Every tool call a session makes is recorded in its own SQLite store (obs.db, separate from conscio.db), searchable later at 0 LLM tokens — so a smaller context window stops meaning lost work. On Claude Code the plugin wires this up automatically; the capture never alters tool output and never blocks a session.

engine.observe(tool="Bash", input_text="ls", output_text="README.md", session_id="s1")
hits = engine.recall_observations("README", session_id="s1")  # FTS5 snippet window
handoff = engine.compress_observations(session_id="s1")       # session → handoff

Recall is session-scoped by default, so a session only ever mines its own trail unless you widen scope.

conscio govern status    # ceiling, obs.db size, capture health, baseline
conscio govern prefix    # measure your stable prefix and where compactions land
conscio govern on        # freeze a baseline + apply the cost-optimal window here
conscio govern report    # current vs baseline, from the host's own usage records
conscio govern off       # restore what you had before

govern report prices both sides per turn rather than as totals — a total against a total mostly measures which side ran longer — and prints — where the baseline froze no figure to compare against, rather than a zero that would render as a 100% saving.

Capture is complete, not truncated: a tool's whole input and output are stored (up to 1 MiB per field), so anything a tool reads or writes — including secrets — can land in obs.db. It is a second copy of what the host already keeps in its session transcripts, with a 30-day retention window. Treat it as one more place to scrub, and lower max_age_days if that matters to you.

Embodiment — the MCP server

conscio-mcp is a hand-rolled, stdlib-only MCP stdio server (newline-delimited JSON-RPC 2.0), so any MCP host can plug into a Conscio instance and consume its cognition live. Zero new dependency; nothing opens a socket.

The tool surface is sized to the model. Three nested surfaces, so raising one never removes a tool:

Precedence is --mode on the CLI, then the persisted choice, then the default. conscio_mode switches at runtime and is present in every surface — in lite it is the only way back out. An unadvertised tool stays callable through tools/call, and tools enabled by flag (act, review, relay) are never filtered — 42 with every flag on.

The base surface is propose-only (perceive / reflect / recall / audit); opt-in --enable-act adds host-executed, ledgered, gated act — Conscio signs and audits the intent, the host pulls the trigger. See the MCP guide.

Live mode — daemon, sensors & Awake Mode

Conscio can run as a living process that perceives the world each cycle and acts only when explicitly awake (R9, default OFF):

from conscio import ConsciousnessEngine, HostSensor
from conscio.daemon import Daemon

engine = ConsciousnessEngine("glm-5.1", storage_path="~/.conscio/live")
engine.wake()                                              # opt in to autonomy (persisted)
Daemon(engine, sensors=[HostSensor()], interval=30).run()  # perceive → reflect → act

conscio-daemon --sensors host --interval 30 runs it standalone (add --awake); conscio awake reaches an already-running daemon rather than waking a second engine of its own. Reference sensors HostSensor / AgentSensor ship as conscio.sensors entry points; write your own SensorAdapter.

Society — shared minds

Same-host instances can share locally-proven skills as data through a host-shared noosphere.db (publish → static-revalidated quarantine → sandboxed trial → promotion), audit each other's action records, and exchange messages over the Liaison mailbox. Engine-free, read-only on the live conscio.db, no inherited trust, no network.

Intercepter

A model asked for 0.15 * 8000 will happily invent a number. Intercepter takes the [INTERCEPT: ...] expressions it emits and evaluates them — a restricted AST walker over arithmetic, comparisons, a fixed set of math functions and bound variables. No eval, no exec, no attribute access, no imports; expressions are length- and depth-capped. LaTeX is converted before parsing, and an equation with no bound variables is solved rather than evaluated.

from conscio.agency.intercepter import Intercepter

itc = Intercepter()
itc.set_variable("rate", 0.15)
itc.process("[INTERCEPT: rate * 8000]").text
# '[INTERCEPT: rate * 8000] -> [RESULT: 1200.0]'

attach_adapter(..., intercept_enabled=True) wires it into the act pipeline as an InterceptionLoop: the model emits a tag, gets the real answer back, and may revise — up to 3 iterations, each one an EventBus record.

Architecture

            reflect  ── passive · advisory · append-only ──┐
                                                              │
  ConsciousnessEngine  (orchestrator · lifecycle · injection) │
   ├─ Witness        InnerMonologue · WorldModel · MetaCognition · GoalGenerator
   ├─ Substrate      ContentStore (FTS5 BM25 + RRF) · VectorBackend (HNSW/sqlite-vec/numpy) ·
   │                 HybridRetriever · EventBus · FilterPipeline
   ├─ Continuity     SessionLifecycle (6-step handoff) · SessionRAG (optional)
   ├─ Metabolism     MetabolicContext · DreamCycle (release→prune→…→distill)
   ├─ Coherence      CoherenceEngine · semantic reconciliation · unmeasured dimensions named
   ├─ Structural     StructuralDistiller (graph → ranked signal; data, not code)
   ├─ Evaluation     evaluate — 5/6-axis rubric
   ├─ Gates          decide · council · loop_gate · delivery_check · investigate
   ├─ Pipelines      acceptance_criteria · verify · continuous_loop ·
   │                 strategic_compact · ledger
   ├─ Diagnostics    context_budget · eval_harness · rules_distill
   ├─ Harness        PromptZones · CheckpointChain · TokenAccount+CPM ·
   │                 FailureGovernor · adaptive max_retries · skeptic skip
   ├─ Adaptive       prompt_complexity (full/compact/minimal) ·
   │                 auto-detect (--model auto) · FallbackAdapter
   ├─ Memory         KnowledgeGraph · Hallways · WingManager · Deduplicator ·
   │                 EntityDetector · EmbeddingProvider · Miner · Migration
   ├─ Observations   obs.db (FTS5, separate) · capture hooks ·
   │                 recall_observations · compress_observations
   ├─ Governor       prefix/landing measurement → cost-optimal window · baseline
   │                 + report, priced per turn from the host's usage records
   ├─ Intercepter    restricted-AST evaluation of [INTERCEPT: ...] tags
   └─ Embodiment     conscio-mcp: JSON-RPC 2.0 over stdio (stdlib only)
                                                              │
            act  ── opt-in agency · audited · gated ◀───────┘
              Skeptic (hostile audit) · TrustMatrix (earned autonomy) ·
              CircuitBreaker (per-goal quarantine) · ActionLedger (append-only)

Subsystem detail and the full public-API reference live on the docs site (docs/, built with mkdocs build --strict).

Any model

Conscio is model-agnostic — it runs on any backend (local Ollama / llama.cpp / LM Studio, any OpenAI-compatible endpoint, or a frontier API). The only thing it needs from a model is its context window: that single number selects the injection mode, and nothing else is hardcoded to a particular model. A known model resolves offline and deterministically; an unknown one is inferred from its name or pinned explicitly:

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

Adaptive prompt complexity. ProbeSuite measures each model's json_fidelity, schema_depth and instruction_depth (5 empirical probes, cached in SQLite), and prompt_complexity picks a prompt tier from the profile:

Auto-detect + fallback chain. --model auto makes the MCP config fixed forever — on boot Conscio calls GET /v1/models, filters out embedding models, tests each chat model with a minimal prompt, and uses the first that responds. The winner is persisted, so the next boot starts instantly. At runtime FallbackAdapter switches to the next model in the chain on a permanent error, timeout or bad response.

Benchmark (local LM Studio, 5 cycles)

Both small models hit 100% JSON validity through Conscio (raw: 60% and 80%). Token cost is 6–10× without prompt caching; with provider caching the stable zone caches at ~0.1×, bringing the effective cost to ~2–3×.

Vector backends

Three backends, auto-detected at startup — no config needed. Priority is HNSW → sqlite-vec → numpy; override with CONSCIO_VEC_BACKEND.

Measured on 37,042 real embeddings, 384-dim, all-MiniLM-L6-v2:

HNSW is 62× faster than numpy and 6× faster than sqlite-vec.

conscio migrate-vectors                  # → sqlite-vec (10×, default target)
conscio migrate-vectors --target hnsw    # → HNSW (50×, direct)

Every path auto-detects the source format, backs up first, verifies search rankings, and writes the target. HNSW writes to a separate hnsw.db, so the original vectors.db is never clobbered. Full guide: docs/MIGRATION.md.

Bench

conscio-bench --adapter mock                          # offline, deterministic
conscio-bench --adapter ollama:qwen3.5:0.8b --cycles 20
conscio-bench --adapter mock --skills 20              # skill-acquisition curve

Backends: mock, ollama:<model>, llamacpp[:<name>], lmstudio:<model>[@<base_url>], openai:<model>[@<base_url>]. Reports probe profile, decode tier, per-tier syntactic validity, Skeptic catch-rate, latency p50, and calibration. Baselines in docs/bench/.

Extending Conscio

Three stable extension points, usable directly or published by a third party and auto-discovered via entry points (conscio.adapters / conscio.sensors / conscio.tools):

# 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. Discover what is installed with conscio plugins.

Testing & data

# House rule: one file per pytest process (the full run OOMs on small machines; CI matches)
for f in tests/test_*.py; do pytest "$f" -q; done
pytest tests/test_agency_act.py -v    # a specific module

SQLite in WAL mode. The engine's storage defaults to ~/.hermes/consciousness/, where conscio.db holds EventBus + ActionLedger + skills, content_store.db holds the ContentStore, and obs.db holds tool observations in a store of its own. Vector backends write to vectors.db (sqlite-vec/numpy) and hnsw.db. Pass storage_path= (or --storage) to move it; the CLI and daemon additionally honour $HERMES_HOME, which the library default does not read. Cross-instance state — the knowledge graph, hallways, vectors, dedup, handoffs, the act sandbox — lives under ~/.conscio/. Always call engine.close() or use the with statement so WAL checkpoints flush.

License

AGPL-3.0-or-later — Neguiolidas / Neguitech