MCP tools reference¶
extremis exposes 10 tools to Claude via the MCP protocol.
memory_remember¶
Append a message or fact to the memory log and write it as an episodic memory.
| Parameter | Type | Default | Description |
|---|---|---|---|
content |
string | required | The text to remember |
role |
string | "user" |
Who said it — "user", "assistant", or "system" |
conversation_id |
string | "default" |
Groups messages for consolidation |
LLM cost: None
memory_recall¶
Retrieve memories relevant to a query using semantic search.
| Parameter | Type | Default | Description |
|---|---|---|---|
query |
string | required | What you're looking for |
limit |
int | 10 |
Max memories to return |
layers |
string | "" |
Comma-separated filter e.g. "semantic,procedural" |
LLM cost: None
memory_report_outcome¶
Apply a reinforcement signal to recalled memories.
| Parameter | Type | Default | Description |
|---|---|---|---|
memory_ids |
string | required | Comma-separated UUIDs from recall results |
success |
bool | required | true = positive signal, false = negative |
weight |
float | 1.0 |
Signal magnitude |
LLM cost: None
memory_remember_now¶
Write directly to a structured memory layer, bypassing the log.
| Parameter | Type | Default | Description |
|---|---|---|---|
content |
string | required | The memory content |
layer |
string | "semantic" |
episodic, semantic, procedural, identity, working |
confidence |
float | 0.9 |
Certainty (0.0–1.0) |
expires_at |
string | "" |
ISO8601 expiry for working memories |
LLM cost: None
memory_consolidate¶
Distil new log entries into structured semantic and procedural memories.
Note
Only processes log entries since the last checkpoint. Does not touch existing structured memories. See memory_compact for reconciling existing memories.
LLM cost: Claude Haiku per conversation batch
memory_compact¶
Reconcile contradictions in existing structured memories.
| Parameter | Type | Default | Description |
|---|---|---|---|
layer |
string | "semantic" |
Which layer to compact |
LLM cost: Claude Haiku per batch of memories
memory_kg_write¶
Add entity, relationship, or attribute to the knowledge graph.
| Parameter | Type | Default | Description |
|---|---|---|---|
operation |
string | required | add_entity, add_relationship, add_attribute |
name |
string | "" |
Entity name (for all operations) |
entity_type |
string | "concept" |
person, org, project, group, concept, other |
from_entity |
string | "" |
Source entity (for relationships) |
to_entity |
string | "" |
Target entity (for relationships) |
rel_type |
string | "" |
Relationship label |
weight |
float | 1.0 |
Confidence |
key |
string | "" |
Attribute key |
value |
string | "" |
Attribute value |
LLM cost: None
memory_kg_query¶
Query the knowledge graph for an entity and its connections.
| Parameter | Type | Default | Description |
|---|---|---|---|
name |
string | required | Entity to look up |
traverse_depth |
int | 0 |
BFS depth (0 = entity + direct connections) |
LLM cost: None
memory_observe¶
Compress recent log entries into priority-tagged observations.
| Parameter | Type | Default | Description |
|---|---|---|---|
conversation_id |
string | "default" |
Which conversation to observe |
Returns entries tagged 🔴 CRITICAL / 🟡 CONTEXT / 🟢 INFO.
LLM cost: None (pure heuristic)
memory_score_attention¶
Score an incoming message 0–100 to decide how much attention to give it.
| Parameter | Type | Default | Description |
|---|---|---|---|
message |
string | required | The incoming message |
sender |
string | "" |
Sender identifier |
channel |
string | "dm" |
dm, group, or broadcast |
owner_ids |
string | "" |
Comma-separated owner IDs (always full attention) |
allowlist |
string | "" |
Comma-separated elevated senders |
ongoing |
bool | false |
Part of an active conversation |
already_answered |
bool | false |
Someone else already replied |
LLM cost: None (pure heuristic)