Search the knowledge graph and store new knowledge.
Query to search for: $ARGUMENTS
Search memory by natural language query using the retrieve machine:
retrieve(action: "context", query: "<query>", limit: 10, expansion_hops: 1)All parameters:
action (required) — "context" for κ-aware ranked retrieval
query (required) — natural language search
limit — max results (default 10)
expansion_hops — 0 = fast/precise, 1 = contextual neighbors, 2 = deep discovery
neighbors_per_node — how many neighbors per hop (default 4; use 8 for deep discovery)
node_type — filter: "semantic", "procedural", "episodic", "temporal", "outcome", "goal"
min_score — similarity threshold (0.0–1.0)
Deep discovery example:
retrieve(action: "context", query: "auth architecture", expansion_hops: 2, neighbors_per_node: 8)Each result node includes:
similarity — how closely the node matches your query (0.0–1.0)
hops — 0 = direct match, 1+ = reached via expansion
via — which edge/node path reached this result (for expansion hits)
causal_context — save this array — you need it for learn_from_outcome
The response includes a topology object:
routing: "fast" (no cycles, proceed normally) or "deliberate" (cycles found — use /graphonomous:deliberate)
max_kappa — cycle complexity (0 = acyclic)
scc_count — number of strongly connected components found
retrieve machine)retrieve(action: "episodic", since: "2026-03-01T00:00:00Z", until: "2026-03-29T23:59:59Z", limit: 20)Time-range filtered retrieval of episodic nodes, sorted by recency. All parameters are optional.
retrieve(action: "procedural", query: "how to deploy the auth service", limit: 10)Semantic search scoped to procedural nodes. Returns nodes with extracted steps from content.
After acting on retrieved context, store what you learned:
act(action: "store_node", content: "<one atomic fact>", node_type: "semantic", confidence: 0.7, source: "conversation")Node types:
semantic — facts, definitions, architecture ("what is?")
procedural — workflows, how-to, recipes ("how to?")
episodic — events, observations, outcomes ("what happened?")
temporal — time-bound observations, monitoring events ("when did X happen?")
outcome — measured results, benchmark scores ("what was the result?")
goal — objectives, targets, intent ("what are we trying to achieve?")
Confidence calibration:
0.9–1.0: Directly observed in code/docs
0.7–0.89: Strong evidence, not directly verified
0.5–0.69: Single source, plausible
0.3–0.49: Indirect evidence, uncertain
0.0–0.29: Speculative
Connect related nodes when it improves retrieval:
act(action: "store_edge", source_id: "<id>", target_id: "<id>", edge_type: "supports", weight: 0.8)Edge types: causal, causes, resolves, supports, contradicts, related, related_to, part_of, follows, supersedes, depends_on, similar_to, derived_from, temporal_before, temporal_after, co_occurs
Weight guidance:
0.8–1.0: Strong, well-evidenced relationship
0.5–0.7: Moderate, reasonable inference
0.2–0.4: Weak, tentative connection
Skipping retrieval before acting (amnesia agent)
Storing kitchen-sink nodes with multiple facts — one atomic fact per node
Discarding causal_context from retrieval — hold it until outcome is resolved
Setting all confidence to 0.9+ — calibrate honestly
Storing duplicates — use query_graph(operation: "similarity_search") first; similarity > 0.90 = duplicate