This page walks through the Graphonomous runtime from process startup to closed-loop learning, goal review, and maintenance.
If you want a practical mental model, think in this sequence:
Boot services
Retrieve context
Act
Store durable knowledge
Learn from outcomes
Review goal coverage
Run consolidation
Graphonomous runs as an OTP application with supervised runtime components:
Store
Embedder
Graph
Retriever
Learner
GoalGraph
Attention
Consolidator
MCP registry/server plumbing
Store opens the SQLite database.
Schema/migrations are applied.
ETS hot cache is rebuilt from durable state.
Core services come online and MCP tools/resources are exposed.
You get:
fast reads (ETS)
durable writes (SQLite)
process-level fault isolation (OTP supervision)
Most integrations begin through MCP over stdio.
Typical first checks:
Runtime health snapshot
Goal snapshot
Lightweight context retrieval for the current task
This gives the agent session orientation before making changes.
Graphonomous retrieval is not only semantic similarity. It combines:
embedding similarity search
graph-neighbor expansion
confidence-aware ranking
topology analysis (SCC/κ routing metadata)
results: ranked candidate context nodes
causal_context: node IDs used later for outcome learning
topology.routing: fast or deliberate
topology.max_kappa: cycle complexity signal
fast: proceed with normal reasoning
deliberate: use deeper deliberation for cyclic/conflicting regions
After retrieval, the agent reasons and executes work. New durable knowledge should be stored in atomic form.
semantic: facts/architecture
procedural: how-to/workflows
episodic: observed events or session outcomes
temporal: time-indexed observations, monitoring events
outcome: empirical results of actions (grounding)
goal: durable intent, objectives, targets
one claim/procedure/event per node
realistic confidence values
source attribution when possible
edges only when relationship quality justifies it
After a meaningful action, Graphonomous can update confidence on the causal nodes that informed that action.
action_id
status (success, partial_success, failure, timeout)
outcome confidence
causal_node_ids (from retrieval causal context)
optional evidence/trace metadata
Learner applies a bounded, learning-rate blend from prior confidence and status-scaled signal, so graph trust adapts over time.
useful nodes get reinforced
weak/wrong assumptions are down-weighted
retrieval quality improves across sessions
For multi-step work, goals provide durable intent and lifecycle management.
proposed -> active -> completed
active -> blocked (if escalation needed)
optional pause/abandon transitions as policy requires
Coverage computes:
coverage_score
uncertainty_score
risk_score
decision: act | learn | escalate
This lets the runtime gate actions based on epistemic readiness, not just availability of any context.
When many goals are in flight, attention can survey and rank what needs focus.
Autonomy modes:
observe (read-only prioritization)
advise (recommend actions)
act (dispatch bounded actions)
Attention combines urgency, coverage state, and topology complexity to prioritize execution.
Consolidation keeps memory quality healthy over time.
The 7-stage pipeline includes:
Confidence decay
Prune weak nodes
Prune weak edges
Strengthen co-activated edges
Merge similar nodes
Promote timescale (fast → medium → slow → glacial)
Generate abstractions from episodic clusters
Operationally, run consolidation:
at session boundaries
after heavy write bursts
periodically during long autonomous loops
A complete interaction often looks like this:
Agent retrieves context for user task.
Runtime returns ranked results + causal context + topology signal.
Agent performs work and returns output.
Agent stores key new semantic/procedural/episodic nodes.
Agent reports outcome against causal nodes.
Goal progress is updated.
Coverage is reviewed before next consequential step.
Consolidation runs to maintain graph quality.
Before ending a productive session:
[ ] Store key durable knowledge
[ ] Report outcomes for consequential actions
[ ] Update linked goals/progress/status
[ ] Trigger or verify consolidation
[ ] Confirm runtime remains healthy
skipping retrieval on non-trivial tasks
losing causal context before outcome reporting
inflating confidence without evidence
storing kitchen-sink nodes
fabricating coverage or outcome signals
ignoring repeated learn/escalate routing
quickstart
architecture
mcp-tools
operations
skills/SKILLS
If you are integrating an autonomous agent, treat this walkthrough as the runtime backbone and the skills docs as strict operating policy.