Shodh Memory

Delete a comment from a todo.

Dismiss/acknowledge a triggered reminder. Call this after you've handled a reminder.

Walk the knowledge graph outward from a named entity: connected entities by hop distance, and the typed relationships between them (Causes, Triggers, DependsOn, custom types, plus co-occurrence), each with strength and source context. Use when you have an entity NAME (person, system, place, ship, ...) and want its connections — recall searches episodic text instead, and trace_lineage follows memory-to-memory causality instead. Name matching is fuzzy (case-insensitive, stems, substrings); the output states which entity actually matched. Use list_entities to browse what exists.

Get synthesized narratives from accumulated facts, clustered by topic with confidence levels and causal chains. Shows what the system has learned, organized into coherent themes.

Delete a specific memory by ID

Rank recent memories by statistical deviation from this user's own rolling baseline (novel entities, unusual entity co-occurrence, untyped-relation share). Each entry carries per-component z-scores and a deterministic explanation of why it deviates. Use to answer 'what has been unusual lately' or to spot weak signals worth investigating — this is deviation scoring against the corpus's own shape, not content search; there is no query. The baseline is the most recent 200 scored episodes (server default); at least 10 must exist before anything is scored.

Survey the causal lineage graph: totals by relation type and source (inferred/confirmed/explicit), average confidence, and the highest-confidence edges with their edge IDs. Use to see what causal structure exists across all memories, or to find edge IDs for validate_causal_link. For the chain around one specific memory, use trace_lineage instead.

List the entities in the knowledge graph, ranked by salience (learned importance), with their types and mention counts. Use to discover what people, systems, places, and concepts the graph tracks — or to find the exact name to pass to explore_entity. For memory contents use recall; this is the graph's cast of characters.

List all stored memories

List all projects with todo counts and status breakdown.

List all pending reminders. Use to check what reminders are scheduled.

List subtasks of a parent todo. Use add_todo with parent_id to create subtasks.

List all comments and activity history for a specific todo.

List or search todos. Supports semantic search via query parameter, or GTD-style filtering. Returns Linear-style formatted output grouped by status.

Get statistics about stored memories

REQUIRED: Call this tool with EVERY user message to surface relevant memories and build conversation history. Pass the user's message as context. This enables: (1) retrieving memories relevant to what the user is asking, (2) building persistent memory of the conversation for future sessions. The system analyzes entities, semantic similarity, and recency to find contextually appropriate memories. Auto-ingest stores the context automatically. USAGE: Always call this FIRST when you receive a user message, passing their message as the context parameter. Surfaced memory bodies are previews (default…

Delete facts matching a content pattern. Use dry_run=true to preview before deleting. Useful for cleaning up garbage facts (e.g., 'relates to' template noise).

Read the FULL content of a specific memory by ID. Use this when you need to see the complete text of a memory that was truncated in search results.

Search memories AND todos using semantic similarity. Returns both relevant memories and matching todos. Use this to find past experiences, decisions, context, or pending work. Modes: 'semantic' (vector similarity), 'associative' (graph traversal), 'temporal' (time-based retrieval), 'hybrid' (combined), 'spatial' (geo-location based), 'mission' (mission context), 'action_outcome' (reward-based learning). Memory bodies are returned as previews (default 500 chars); when truncated the output carries an explicit marker with real lengths and a read_memory hint — a preview without a marker is complet…

Find memories by tags. Returns memories matching ANY of the provided tags. Useful for finding memories by category (e.g., 'tool:Edit', 'file:src/main.rs', 'source:hook', 'error', 'session-summary'). Memory bodies are returned as previews (default 500 chars); when truncated the output carries an explicit marker with real lengths and a read_memory hint — a preview without a marker is complete. Pass full_content:true to get full bodies inline.

Give Hebbian feedback on memories after using them: outcome 'helpful' boosts their importance and strengthens their associations so they surface more readily, 'misleading' decays them, 'neutral' just records the access. Call after completing a task with the memory IDs recall gave you — this is how retrieval learns from outcomes instead of only from similarity.

Store a memory for future recall. Use this to remember important information, decisions, user preferences, project context, or anything you want to recall later.

Move a todo up or down within its status group. Use to prioritize tasks manually.

Repair vector index by re-indexing orphaned memories. Use this when verify_index shows unhealthy status. Returns count of repaired memories.

Reset the token counter for a new session. Call this when starting a new conversation or after context has been compressed/summarized.

Search the distilled semantic facts the system has consolidated out of episodic memories — stable knowledge like preferences, capabilities, relationships, and procedures, each with a confidence score and supporting-memory count. Pass query for keyword search, entity for facts about one entity, or neither to list recent facts. Use for 'what does the system KNOW about X' — recall searches raw episodic memories instead, and fact_narratives returns topic-clustered summaries instead of individual facts.

Get a consolidated digest of the current session: timestamps, token usage, memories created/recalled, tools used with counts, entities extracted, topic changes, and consolidation events. Use after context compression or at session milestones.

Show what you worked on across recent sessions. Returns session summaries with entities, memory stats, and timestamps. Use group_by_project to detect cross-session project continuity via entity overlap.

Set a reminder for the future. Triggers on time (at specific time or after duration) or context match (when keywords appear in conversation). Reminders will surface automatically when conditions are met.

Get statistics about your todos - counts by status, overdue items, etc.

Get MCP pipeline token throughput for this session. Tracks tokens flowing through shodh memory tools only — NOT the AI context window. Use for internal diagnostics, not context window health.

Trace the causal chain of a memory: what led to it (direction=backward, the default) or what it went on to cause (forward). Follows typed causal edges between memories (Caused, ResolvedBy, InformedBy, SupersededBy, TriggeredBy) and reports the root cause — the oldest ancestor — when tracing backward. Use this, not recall, to answer 'why did X happen' or 'what did X cause' for a memory whose ID you have (from recall/read_memory output). For exploring around a named entity instead of a memory, use explore_entity.

Update a todo's properties. Use short ID prefix (e.g., SHO-1a2b) or full ID.

Update an existing comment on a todo.

Confirm or reject a causal edge by its edge ID (shown by trace_lineage and list_causal_edges). Confirming an inferred edge raises it to full confidence and strengthens the knowledge-graph connections between the two memories' entities; rejecting deletes the edge. Use when the user or the evidence settles whether an inferred causal link is real — confirmed chains make root-cause tracing trustworthy.

Verify vector index integrity - diagnose orphaned memories that are stored but not searchable. Returns health status and count of orphaned memories.