This page answers the most common questions about Graphonomous, its MCP interface, and operating model.
Graphonomous is a continual-learning memory engine for AI agents. It stores knowledge as a graph, supports confidence-updating feedback loops, and exposes all capabilities through MCP tools/resources.
No. Graphonomous does not retrain model weights. Learning happens in the memory graph by storing knowledge and updating confidence based on outcomes.
It gives agents durable, evolving memory and better decision quality over time by combining:
semantic retrieval
graph neighborhood expansion
causal outcome feedback
goal tracking
topology-aware routing (fast vs deliberate)
Use node types intentionally (6 types):
semantic — facts, architecture, definitions
procedural — workflows and how-to steps
episodic — events and observations
temporal — time-indexed observations, monitoring events
outcome — empirical results of actions (grounding)
goal — durable intent, objectives, targets
16 edge types (including 2 legacy aliases):
causes, resolves — causal attribution
supports, contradicts — evidential
related_to, similar_to — topical affinity
part_of, follows, supersedes, depends_on — structural
temporal_before, temporal_after, co_occurs — temporal ordering
derived_from — provenance
Legacy aliases: causal, related (backward-compatible)
Default edge weight is 0.3. Edges should be added only when they improve retrieval quality or provenance clarity.
retrieve_context generally does:
similarity retrieval
graph expansion
ranking with confidence-aware scoring
topology analysis (SCC/κ-aware metadata)
Results include a causal_context array for later feedback via learn_from_outcome.
causal_context and why does it matter?causal_context is the list of node IDs that informed the current action. You should pass those IDs into learn_from_outcome so Graphonomous can update confidence on the actual causal nodes.
learn_from_outcome supports:
success
partial_success
failure
timeout
Use timeout when an action did not complete in time (instead of incorrectly marking it as failure).
Graphonomous includes durable GoalGraph operations:
create/list/get/update/delete goals
transition status
set progress
link/unlink evidence nodes
run epistemic review (review_goal)
Typical statuses include proposed, active, blocked, completed, and abandoned.
review_goal evaluates whether current knowledge is sufficient to proceed. It returns decision-oriented signals such as:
act (enough coverage)
learn (need more context)
escalate (insufficient/too risky)
Graphonomous analyzes retrieved subgraphs for cycles and complexity. If topology is simple, routing is fast. If cyclic complexity is higher, routing may indicate deliberate, signaling deeper reasoning is safer.
Consolidation is a 7-stage periodic memory maintenance pipeline:
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
Use run_consolidation to trigger/inspect cycles manually when needed.
Yes. It is designed for local operation with practical defaults and fallback modes. A common setup uses a local SQLite DB path and a lightweight embedder backend configuration for reliability.
Typical options:
run as MCP server over stdio
run one-shot traversal/scan mode
run watch mode for ongoing ingestion
See quickstart and runtime docs pages for exact commands and settings.
For non-trivial work, use this loop:
retrieve context
reason and act
store durable knowledge
report outcomes
maintain via consolidation
This is the recommended operating pattern for robust continual learning.
See the MCP tools reference pages in this docs site, especially:
retrieval and graph query
node/edge storage
outcome learning
goal management and review
topology analysis and deliberation
attention and consolidation
Yes—short version:
retrieve first for meaningful tasks
store atomic nodes (avoid “kitchen-sink” nodes)
keep confidence calibrated to evidence quality
preserve causal IDs for feedback
avoid fabricated provenance/signals
review goals before consequential actions
consolidate periodically
If you want a deeper operational playbook, start with the docs quickstart, then architecture, then the workflows and anti-patterns sections.