Data model

Schema · SQLite memory

Field-by-field reference for the SQLite memory tables — the agents’ persistent knowledge graph.

Under review — agent memory may move to Convex

The agent-memory layer is likely to use Convex (native vector search, full-text search and the @convex-dev/agent component) rather than a separate SQLite service — Convex is already self-hosted, reactive and backed up. SQLite stays only as the embedded Convex backend store; pgvector / Qdrant is the graduation path at scale. The schema below is kept for reference and is revisited when the agentic layer is built.

SQLite stores the agents’ persistent knowledge as a lightweight knowledge graph around one central memory_items table — not raw market data. Every conclusion keeps its evidence; revisions are immutable.

Memory tables

memory_items

Field	Type	Description
`id`	string	Memory ID
`memory_type`	string	Episodic, semantic, procedural, experimental, human
`entity_type`	string	Token, opportunity, strategy, trade
`entity_id`	string, optional	Related operational entity
`title`	string	Memory title
`summary`	text	Concise knowledge
`content`	JSON	Structured memory content
`importance`	number	Importance score
`confidence`	number	Confidence score
`valid_from`	datetime, optional	Beginning of validity
`valid_until`	datetime, optional	End of validity
`status`	string	Active, superseded, invalidated, archived
`source_type`	string	Trade, experiment, human, agent
`source_id`	string, optional	Source entity
`created_by`	string	Human, agent or system
`created_at`	datetime	Creation date
`updated_at`	datetime	Last update
`last_accessed_at`	datetime, optional	Last retrieval
`access_count`	integer	Retrieval count

memory_relations

Field	Type	Description
`id`	string	Relation ID
`source_memory_id`	string	Source memory
`target_memory_id`	string	Target memory
`relation_type`	string	Supports, contradicts, caused, similar_to
`strength`	number	Relation strength
`confidence`	number	Confidence
`created_by`	string	Human, agent or system
`created_at`	datetime	Creation date

memory_evidence

Field	Type	Description
`id`	string	Evidence ID
`memory_id`	string	Related memory
`evidence_type`	string	Market data, trade, report, human note
`source_reference`	string	Operational or analytical reference
`direction`	string	Support, contradict, neutral
`weight`	number	Evidence weight
`observed_at`	datetime, optional	Evidence event time
`summary`	text	Evidence summary
`data`	JSON, optional	Structured evidence
`created_at`	datetime	Creation date

memory_embeddings

Field	Type	Description
`memory_id`	string	Related memory
`embedding_model`	string	Model used
`embedding_version`	string	Model/version identifier
`embedding`	vector/blob	Vector representation
`content_hash`	string	Detects outdated embedding
`created_at`	datetime	Creation date

memory_feedback

Field	Type	Description
`id`	string	Feedback ID
`memory_id`	string	Retrieved memory
`workflow_run_id`	string, optional	Workflow using memory
`feedback_type`	string	Useful, irrelevant, incorrect, outdated
`score`	number	Feedback score
`note`	text, optional	Explanation
`created_by`	string	Human or agent
`created_at`	datetime	Creation date

memory_revisions

Field	Type	Description
`id`	string	Revision ID
`memory_id`	string	Related memory
`revision_number`	integer	Revision
`previous_content`	JSON	Previous state
`change_reason`	text	Reason for change
`changed_by`	string	Human, agent or system
`created_at`	datetime	Revision date

memory_compaction_runs

Field	Type	Description
`id`	string	Compaction run ID
`status`	string	Running, complete, failed
`scope`	string	Token, strategy, period, global
`source_memory_count`	integer	Memories processed
`created_memory_count`	integer	Consolidated memories created
`superseded_memory_count`	integer	Memories superseded
`summary`	text	Compaction summary
`started_at`	datetime	Start time
`completed_at`	datetime, optional	Completion time

Structured first, vectors last

Retrieval prefers structured fields, then FTS5, then sqlite-vec similarity. memory_embeddings is optional and never replaces structured queries.