Skip to content

MCP Tools

All tools are available when engram is connected as an MCP server (engram-mcp).

Tool Reference

Tool Description
engram_remember Store memory with type, priority, tags, namespace
engram_recall Search episodic memories (compact format by default)
engram_think Reason across episodic + semantic memory via LLM
engram_status Show memory statistics
engram_get_memory Retrieve full memory content by ID or prefix
engram_timeline Chronological context around a memory
engram_add_entity Add entity node to knowledge graph
engram_add_relation Add relationship edge between entities
engram_query_graph Query knowledge graph
engram_ingest Dual ingest: extract entities + store memories
engram_meeting_ledger Record structured meeting (decisions, action items)
engram_feedback Record positive/negative feedback on memories
engram_auto_feedback Auto-detect feedback from conversation context
engram_cleanup Delete all expired memories
engram_cleanup_dedup Deduplicate similar memories by cosine similarity
engram_summarize Summarize recent N memories via LLM
engram_session_start Begin new conversation session
engram_session_end End active session with optional summary
engram_session_summary Get summary of completed session
engram_session_context Retrieve memories from active session
engram_ask Smart query — auto-routes to recall or think

Tool Details

engram_remember

Store a memory with optional metadata.

Parameters:

Parameter Type Description
content string The memory content
memory_type string fact, decision, preference, todo, error, context, workflow, meeting_ledger
priority int 1-10, default 5
tags list[str] Optional tags for filtering
namespace string Tenant namespace
expires string Expiry: 2h, 1d, 7d
topic_key string Unique key for upsert (replaces existing)

engram_recall

Search episodic memories by semantic similarity.

Parameters:

Parameter Type Description
query string Search query
limit int Max results, default 5
memory_type string Optional type filter
tags list[str] Optional tag filter
resolve_entities bool Resolve pronoun/entity references
resolve_temporal bool Resolve date/time expressions

engram_think

Run LLM reasoning across all memory (episodic + semantic + federated).

Parameters:

Parameter Type Description
question string The question to reason about

engram_ingest

Extract entities from messages and store as memories. This is the primary tool for bulk ingestion — it runs entity extraction first and only stores messages that contain meaningful entities (entity-gated).

Parameters:

Parameter Type Description
messages list List of {role, content} message objects

engram_meeting_ledger

Record a structured meeting with decisions and action items.

Parameters:

Parameter Type Description
title string Meeting title
decisions list[str] Decisions made
action_items list[str] Tasks to follow up
attendees list[str] Participants
topics list[str] Topics discussed

engram_feedback

Record feedback to adjust memory confidence scores.

  • Positive: +0.15 confidence
  • Negative: -0.2 confidence
  • 3x negative + low confidence: auto-delete

Parameters:

Parameter Type Description
memory_id string Memory ID
feedback string positive or negative

Session Tools

Session tools track context within a conversation:

engram_session_start  →  engram_session_context (during)  →  engram_session_end
                                                              engram_session_summary

Sessions allow the agent to retrieve only the memories from the current conversation, separate from the full memory store.

# Start of session
engram_session_start()
engram_recall("relevant context for current task")

# During session
engram_remember("important decision: use Redis for caching")
engram_add_entity("Redis", type="Technology")

# End of session
engram_session_end(summary="Implemented Redis caching layer")