Purpose: Teach any LLM connected to Graphonomous how to use its MCP machines correctly, idiomatically, and in the right sequence. These reference docs mirror the skills in the ampersand-plugins Claude Code plugin. For Claude Code users, install the plugin directly:
claude plugin add c-u-l8er/ampersand-plugins
Graphonomous is a continual learning engine exposed as 5 MCP machines. It maintains a self-evolving knowledge graph of semantic, procedural, episodic, temporal, outcome, and goal nodes connected by typed, weighted edges. Your job as an LLM agent is to read from and write to this graph as you work, creating a durable memory that improves over time.
Every interaction should follow this rhythm:
retrieve → route → act → learn → consolidate
"What do I know?" → "What should I do?" → "Do it" → "Did it work?" → "Clean up"Graphonomous v0.4 exposes 5 loop-phase machines instead of 29 individual tools. Each machine accepts an action parameter:
| Machine | Phase | Actions |
|---|---|---|
retrieve | "What do I know?" | context, episodic, procedural, coverage, trace_evidence, frontier |
route | "What should I do?" | topology, deliberate, attention_survey, attention_cycle, review_goal |
act | "Do it" | store_node, store_edge, delete_node, manage_edge, manage_goal, belief_revise, forget_node, forget_policy, gdpr_erase |
learn | "Did it work?" | from_outcome, from_feedback, detect_novelty, from_interaction, contradictions |
consolidate | "Clean up" | run, stats, query, traverse |
| File | Skill | What It Teaches |
|---|---|---|
| bootstrap.md | /graphonomous:bootstrap | Session initialization — retrieve context, check goals, survey attention |
| retrieve.md | /graphonomous:retrieve | The foundational read loop — κ-aware ranked retrieval + store new knowledge |
| store.md | /graphonomous:store | Dedicated write path — atomic nodes, confidence calibration, edges |
| learn.md | /graphonomous:learn | Closed-loop learning — outcome, feedback, novelty, contradictions |
| deliberate.md | /graphonomous:deliberate | κ-aware topology analysis and cyclic reasoning |
| consolidate.md | /graphonomous:consolidate | Memory maintenance — 7-stage pipeline, stats, query, traverse |
| goals.md | /graphonomous:goals | Durable intent tracking — goal CRUD and lifecycle |
| belief.md | /graphonomous:belief | AGM-style belief revision — expand, revise, contract |
| forgetting.md | /graphonomous:forgetting | Structured removal — soft, hard, cascade, GDPR, policy pruning |
| epistemic-frontier.md | /graphonomous:epistemic-frontier | Uncertainty-guided investigation — Wilson intervals, information gain |
| trace-evidence-path.md | /graphonomous:trace-evidence-path | Evidence provenance — weighted Dijkstra, Yen's K-shortest paths |
| attention.md | /graphonomous:attention | Autonomous focus — survey, triage, dispatch |
| review.md | /graphonomous:review | Coverage evaluation — act/learn/escalate routing |
| inspect.md | /graphonomous:inspect | Read-only graph inspection — list, get, edges, search, traverse |
| graph-health.md | /graphonomous:graph-health | Combined diagnostics — weak nodes, orphans, staleness |
| workflows.md | /graphonomous:workflows | End-to-end recipes — cold start, debug, Ralph loop, handoff |
| sync.md | /graphonomous:sync | Batch filesystem ingest to knowledge graph |
| watch.md | /graphonomous:watch | Continuous filesystem monitoring with change detection |
| URI | What It Returns |
|---|---|
graphonomous://runtime/health | Runtime health: node/edge counts, consolidator state, uptime |
graphonomous://goals/snapshot | Current GoalGraph snapshot: all goals with status/progress |
graphonomous://graph/node/{id} | Individual node details + connected edges |
graphonomous://graph/recent | Recently added/accessed nodes, sorted by recency |
graphonomous://consolidation/log | Consolidator state + orchestrator plasticity metrics |
| Type | Store When You Learn... | Examples |
|---|---|---|
| semantic | A fact, definition, or architectural truth | "Module X is responsible for Y", "The API uses JWT auth" |
| procedural | A procedure, workflow, or recipe | "To deploy: run X then Y then Z" |
| episodic | Something that happened — an event or observation | "Build failed due to missing dep" |
| temporal | A time-bound observation or monitoring event | "CPU spike at 14:30 during load test" |
| outcome | A measured result or benchmark score | "Latency dropped 40% after caching change" |
| goal | An objective, target, or intent | "Need to migrate auth to new compliance standard" |
Rule of thumb: "what is?" → semantic. "how to?" → procedural. "what happened?" → episodic.
| Type | Meaning | Example |
|---|---|---|
causal / causes | A causes or drives B | "Config change → behavior change" |
supports | A provides evidence for B | "Test result → hypothesis" |
contradicts | A conflicts with B | "Doc says X, but code does Y" |
related / related_to | Thematically connected | "Auth module ↔ User module" |
derived_from | Extracted or derived from | "Summary ← source document" |
supersedes | Replaced by newer version | "Old preference → updated preference" |
resolves | A resolves issue B | "Fix → bug report" |
part_of | A is a component of B | "Function → module" |
follows | A comes after B | "Step 2 → Step 1" |
depends_on | A requires B | "Feature → dependency" |
similar_to | A resembles B | "Pattern A ↔ Pattern B" |
temporal_before / temporal_after | Temporal ordering | "Event 1 before Event 2" |
co_occurs | A and B happen together | "Error X co-occurs with config Y" |
| Range | Meaning | When to Use |
|---|---|---|
| 0.9–1.0 | Verified fact, directly observed | Copied from code/docs, confirmed by test |
| 0.7–0.89 | Strong evidence, not directly verified | Multiple consistent sources |
| 0.5–0.69 | Moderate, reasonable inference | Single source, plausible |
| 0.3–0.49 | Low confidence, uncertain | Indirect evidence, may be outdated |
| 0.0–0.29 | Speculative or likely wrong | Guess, contradicted by other evidence |
Default is 0.5. Always adjust based on evidence quality.
Retrieve prior context:
`` retrieve(action: "context", query: "summary of prior work on <topic>") ``
Check active goals:
`` act(action: "manage_goal", operation: "list_goals", filters: {"status": "active"}) ``
Survey attention:
`` route(action: "attention_survey", include_idle: false) ``
Proceed with the user's request, using retrieved context.
Store new knowledge: act(action: "store_node", ...)
Report outcomes: learn(action: "from_outcome", ...)
Update goal progress: act(action: "manage_goal", operation: "set_progress", ...)
Consolidate: consolidate(action: "run")
For Claude Code users: Install the plugin — claude plugin add c-u-l8er/ampersand-plugins — and skills are available as /graphonomous:<skill> commands.
For other agents/system prompts: Include SKILLS.md first, then whichever skill files are relevant. Minimum viable context: this file alone.
For agent bootstrapping: Use workflows.md for end-to-end recipes.
For repository wiring: These skills are referenced by AGENTS.md, CLAUDE.md, and README.md in the graphonomous repo.