Database Design

The Light-Fabric utilizes a robust PostgreSQL schema to manage the entire lifecycle of agentic workflows, skills, agent execution, channels, and the biomimetic Hindsight memory system. The schema is organized into five logical layers:

1. Workflow Engine

These tables manage the definition and execution of long-running agentic workflows.

wf_definition_t

Stores the Agentic Workflow DSL (YAML) that defines the high-level orchestration logic.

process_info_t & task_info_t

Manage the runtime state of workflow instances (processes) and individual steps (tasks). They include input_data, context_data, and error_info to provide a resilient "scratchpad" for intermediate variables.

worklist_t & worklist_asst_t

Manage task assignments and visibility for human-in-the-loop interactions.


2. Agentic Core (The "Brain & Skills")

These tables define the identity, expertise, and capabilities of individual agents.

agent_definition_t

Defines the agent's persona, product profile, model policy, default execution profile, data boundary, and runtime limits. Enterprise, coding, and personal-assistant definitions share this table; a definition does not own a process or sandbox.

skill_t

Stores the "Expertise" of an agent in Markdown format. Skills are hierarchical and versioned.

tool_t & tool_param_t

The "Hands" of the agent. Defines governed REST/MCP capabilities and typed execution metadata. A tool row or script field is not authority to execute code; untrusted executable assets require an immutable reviewed package and an approved runner sandbox. The target model adds or derives a stable internal tool reference and records a server-owned gateway, runner, workflow, or fixed-service placement, model alias, schema digest, effect class, and dispatch-policy binding. Existing rows with proven current gateway-config linkage migrate as gateway. Ambiguous legacy script rows remain undisclosed until reviewed.

agent_skill_t & skill_tool_t

Maps agents to skills and skills to tools, implementing the Progressive Disclosure pattern where agents only see the tools required for their current skill context.

skill_package_t (proposed)

References immutable signed/scanned skill assets for coding, personal, or external runtime adapters. Package bytes live in object storage. The row binds digest, provenance, entrypoint, compatible profiles, required capabilities, review, revocation, and retention.


3. Hindsight Memory System

A biomimetic memory architecture that transitions from flat logs to structured "atoms of thought."

agent_memory_bank_t

Profiles for memory banks, defining the "Personality and Disposition" (e.g., skepticism, empathy) of the memory layer.

agent_memory_unit_t

The individual "Atoms" of memory. Each unit contains content and a vector embedding (384-dim) for semantic retrieval.

A Knowledge Graph layer that resolves entities and causal/semantic relationships between memory units.


4. Agent Session And Execution

agent_session_t, agent_turn_t, agent_action_attempt_t, and agent_approval_t (proposed)

Store authenticated session ownership, durable FIFO turns, effectful action attempts, policy snapshots, budgets, bound single-use approvals, reconciliation, and terminal results. Runner-backed actions reference the shared execution_attempt_t, but agent-domain fields remain owned by light-agent. Post-approval dispatch creates a new numbered agent action and a new common execution attempt; it never reopens the pre-approval attempt.

Shared runner execution records (proposed)

runner_scheduling_request_t, execution_attempt_t, execution_session_t, execution_session_cleanup_request_t, and execution_input_t store origin-neutral capacity, fencing, normalized results, bounded session cleanup, immutable staged inputs, and session state/version/fencing. Reused sessions can enter a bounded IDLE_APPROVAL_HOLD with hold ID/expiry, cost policy, and checkpoint/patch evidence while the action lease and credentials are gone. Zero active attempts does not imply cleanup for a valid held session; the hold cannot extend idle/max expiry or override close/revocation. Origin services retain their own workflow or agent state. A terminal-attempt transaction emits an identifiers-only wakeup; origins query the authoritative row and use startup/periodic catch-up rather than treating notification delivery as durable state.

agent_session_event_t (proposed)

An append-only authoritative ledger for accepted user messages, model results, actions, approvals, and terminal turn events. It is the source used to rebuild conversation context after a projection conflict.

agent_session_history_t

The materialized transcript used for active conversation context, linked to the session's Hindsight memory bank. It is not the authoritative proof of an effectful tool action. Agent action attempts and append-only session events survive history-version conflicts and can rebuild this projection. See Light-Agent Execution.

5. Personal Assistant Channels And Triggers

agent_channel_binding_t and agent_channel_delivery_t (proposed)

Bind a verified messaging identity to a principal and agent, and deduplicate inbound events/outbound responses. They store credential references rather than channel secrets and do not replace agent turns.

agent_trigger_t (proposed)

Stores scheduled or connector-triggered turn policy, including timezone, quiet hours, rate/cost limits, notification destination, idempotency, and maximum delay.

See Light-Agent Execution and the dedicated Light-Agent runtime implementation plan for constraints and phased rollout.


DDL Specification

-- Workflow Definitions: Stores the Agentic Workflow JSON
CREATE TABLE wf_definition_t (
    host_id             UUID NOT NULL,
    wf_def_id           UUID NOT NULL,
    namespace           VARCHAR(126) NOT NULL,
    name                VARCHAR(126) NOT NULL,
    version             VARCHAR(20) NOT NULL,
    definition          TEXT NOT NULL, -- The Agentic Workflow DSL in YAML
    aggregate_version    BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT TRUE,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, wf_def_id),
    UNIQUE(host_id, namespace, name, version)
);

CREATE TABLE worklist_t (
  host_id              UUID NOT NULL,
  assignee_id          VARCHAR(126) NOT NULL,
  category_id          VARCHAR(126) DEFAULT '(all)' NOT NULL,
  status_code          VARCHAR(10) DEFAULT 'Active' NOT NULL,
  app_id               VARCHAR(512) DEFAULT 'global' NOT NULL,
  aggregate_version    BIGINT DEFAULT 1 NOT NULL,
  active               BOOLEAN NOT NULL DEFAULT TRUE,
  update_user          VARCHAR (255) DEFAULT SESSION_USER NOT NULL,
  update_ts            TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP NOT NULL,
  PRIMARY KEY(host_id, assignee_id, category_id)
);

CREATE TABLE worklist_column_t (
  host_id               UUID NOT NULL,
  assignee_id           VARCHAR(126) NOT NULL,
  category_id           VARCHAR(126) DEFAULT '(all)' NOT NULL,
  sequence_id           INTEGER NOT NULL,
  column_id             VARCHAR(126) NOT NULL,
  aggregate_version     BIGINT DEFAULT 1 NOT NULL,
  active                BOOLEAN DEFAULT TRUE,
  update_ts             TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
  update_user           VARCHAR(126) DEFAULT SESSION_USER,
  PRIMARY KEY(host_id, assignee_id, category_id, sequence_id),
  FOREIGN KEY(host_id, assignee_id, category_id) REFERENCES worklist_t(host_id, assignee_id, category_id) ON DELETE CASCADE
);

CREATE TABLE process_info_t (
  host_id                    UUID NOT NULL,
  process_id                 UUID NOT NULL, -- generated uuid
  wf_def_id                  UUID NOT NULL, -- workflow definition id
  wf_instance_id             VARCHAR(126)       NOT NULL, -- workflow intance id
  app_id                     VARCHAR(512)       NOT NULL, -- application id
  process_type               VARCHAR(126)      NOT NULL,
  status_code                CHAR(1)            NOT NULL, -- process status code 'A', 'C'
  started_ts                 TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP NOT NULL,
  ex_trigger_ts              TIMESTAMP WITH TIME ZONE          NOT NULL,
  custom_status_code         VARCHAR(126),
  completed_ts               TIMESTAMP WITH TIME ZONE,
  result_code                VARCHAR(126),
  source_id                  VARCHAR(126),
  branch_code                VARCHAR(126),
  rr_code                    VARCHAR(126),
  party_id                   VARCHAR(126),
  party_name                 VARCHAR(126),
  counter_party_id           VARCHAR(126),
  counter_party_name         VARCHAR(126),
  txn_id                     VARCHAR(126),
  txn_name                   VARCHAR(126),
  product_id                 VARCHAR(126),
  product_name               VARCHAR(126),
  product_type               VARCHAR(126),
  group_name                 VARCHAR(126),
  subgroup_name              VARCHAR(126),
  event_start_ts             TIMESTAMP WITH TIME ZONE,
  event_end_ts               TIMESTAMP WITH TIME ZONE,
  event_other_ts             TIMESTAMP WITH TIME ZONE,
  event_other                VARCHAR(126),
  risk                       NUMERIC,
  risk_scale                 INTEGER,
  price                      NUMERIC,
  price_scale                INTEGER, -- Scale (number of digits to the right of the decimal) of the risk column. NULL implies zero
  product_qy                 NUMERIC,
  currency_code              CHAR(3),
  ex_ref_id                  VARCHAR(126),
  ex_ref_code                VARCHAR(126),
  product_qy_scale           INTEGER,
  parent_process_id          VARCHAR(22),
  deadline_ts                TIMESTAMP WITH TIME ZONE,
  parent_group_id            NUMERIC,
  process_subtype_code       VARCHAR(126),
  owning_group_name          VARCHAR(126), -- Name of the group that owns the process
  input_data                 JSONB,        -- The initial data that triggered the workflow
  context_data               JSONB,        -- The runtime "scratchpad" for intermediate variables
  error_info                 TEXT,         -- Detailed error or stack trace if the process fails
  aggregate_version   BIGINT DEFAULT 1 NOT NULL,
  active              BOOLEAN DEFAULT TRUE,
  update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
  update_user         VARCHAR(126) DEFAULT SESSION_USER,
  PRIMARY KEY(host_id, process_id),
  FOREIGN KEY(host_id, wf_def_id) REFERENCES wf_definition_t(host_id, wf_def_id) ON DELETE CASCADE
);

CREATE TABLE task_info_t
(
    host_id             UUID NOT NULL,
    task_id             UUID NOT NULL,
    task_type           VARCHAR(126) NOT NULL,
    process_id          UUID NOT NULL,
    wf_instance_id      VARCHAR(126) NOT NULL,
    wf_task_id          VARCHAR(126) NOT NULL,
    status_code         CHAR(1)       NOT NULL, -- U, A, C
    started_ts          TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP NOT NULL,
    locked              CHAR(1)       NOT NULL,
    priority            INTEGER        NOT NULL,
    completed_ts        TIMESTAMP WITH TIME ZONE      NULL,
    completed_user      VARCHAR(126)     NULL,
    result_code         VARCHAR(126)     NULL,
    locking_user        VARCHAR(126)     NULL,
    locking_role        VARCHAR(126)     NULL,
    deadline_ts         TIMESTAMP WITH TIME ZONE      NULL,
    lock_group          VARCHAR(126)     NULL,
    task_input          JSONB,           -- Specific data passed to the task
    task_output         JSONB,           -- Result returned by the task action
    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT TRUE,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, task_id),
    FOREIGN KEY (host_id, process_id) REFERENCES process_info_t(host_id, process_id) ON DELETE CASCADE
);

CREATE TABLE task_asst_t
(
    host_id             UUID NOT NULL,
    task_asst_id         UUID NOT NULL,
    task_id              UUID NOT NULL,
    assigned_ts          TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP NOT NULL,
    assignee_id          VARCHAR(126) NOT NULL,
    reason_code          VARCHAR(126) NOT NULL,
    unassigned_ts        TIMESTAMP WITH TIME ZONE      NULL,
    unassigned_reason    VARCHAR(126)     NULL,
    category_code        VARCHAR(126)     NULL,
    aggregate_version    BIGINT DEFAULT 1 NOT NULL,
    active               BOOLEAN DEFAULT TRUE,
    update_ts            TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user          VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, task_asst_id),
    FOREIGN KEY(host_id, task_id) REFERENCES task_info_t(host_id, task_id) ON DELETE CASCADE
);

CREATE TABLE audit_log_t
(
    host_id             UUID NOT NULL,
    audit_log_id        UUID NOT NULL,
    source_type_id      VARCHAR(126)      NULL,
    correlation_id      VARCHAR(126)      NULL,
    user_id             VARCHAR(126)     NULL,
    event_ts            TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP NOT NULL,
    success             CHAR(1)           NULL,
    message0            VARCHAR(126)     NULL,
    message1            VARCHAR(126)     NULL,
    message2            VARCHAR(126)     NULL,
    message3            VARCHAR(126)     NULL,
    message             VARCHAR(500)     NULL,
    user_comment        VARCHAR(500)     NULL,
    PRIMARY KEY(host_id, audit_log_id)
);

CREATE INDEX audit_log_idx1 ON audit_log_t (source_type_id, correlation_id, event_ts, user_id);

-- Agent Definitions: Stores the "Brain" configuration
CREATE TABLE agent_definition_t (
    host_id             UUID NOT NULL,
    agent_def_id        UUID NOT NULL,
    agent_name          VARCHAR(126) NOT NULL,
    model_provider      VARCHAR(64) NOT NULL,  -- 'openai', 'anthropic', etc.
    model_name          VARCHAR(126) NOT NULL, -- 'gpt-4o', 'claude-3-5-sonnet'
    api_key_ref         VARCHAR(126),          -- Reference to Secret Manager key
    temperature         NUMERIC(3,2) DEFAULT 0.7,
    max_tokens          INTEGER,               -- max number of tokens can be used
    aggregate_version    BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT TRUE,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, agent_def_id),
    UNIQUE(host_id, agent_name)
);


-- Skills: Stores Instructions and Domain Knowledge (The "Expertise")
-- Note: Use entity_tag_t and entity_category_t with entity_type = 'skill'
-- for flat tagging and hierarchical folder structure of skills.
CREATE TABLE skill_t (
    host_id             UUID NOT NULL,
    skill_id            UUID NOT NULL,
    parent_skill_id     UUID,                  -- Self-reference for Hierarchy
    name                VARCHAR(126) NOT NULL,
    description         VARCHAR(500),          -- High-level description for the initial LLM prompt
    content_markdown    TEXT NOT NULL,         -- The actual instructions/prompts

    description_embedding VECTOR(384),          -- For semantic lookup/discovery
    version             VARCHAR(20) DEFAULT '1.0.0',
    aggregate_version    BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, skill_id),
    FOREIGN KEY(host_id, parent_skill_id) REFERENCES skill_t(host_id, skill_id)
);

CREATE INDEX idx_skill_active ON skill_t(active);
CREATE INDEX idx_skill_name ON skill_t(name);

-- Tools: Stores Executable Functions (The "Hands")
CREATE TABLE tool_t (
    host_id             UUID NOT NULL,
    tool_id             UUID NOT NULL,
    name                VARCHAR(126) NOT NULL,
    description         TEXT NOT NULL,         -- Instructions for LLM on when/how to use this tool

    -- Implementation specifics
    implementation_type VARCHAR(50),           -- 'java', 'mcp_server', 'rest', 'python', 'javascript'
    implementation_class VARCHAR(500),         -- FQCN if 'java'
    mcp_server_name      VARCHAR(126),         -- MCP server name if 'mcp_server'
    api_endpoint        VARCHAR(1024),         -- URL if 'rest'
    api_method          VARCHAR(10),           -- HTTP Method if 'rest'
    endpoint_id         UUID,                  -- Reference to fine-grained auth endpoint
    script_content      TEXT,                  -- Source code if 'python'/'javascript'
    response_schema     JSONB,                 -- Strict output schema for tool results

    description_embedding VECTOR(384),          -- For semantic lookup/discovery
    version             VARCHAR(20) DEFAULT '1.0.0',
    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, tool_id),
    FOREIGN KEY(host_id, endpoint_id) REFERENCES api_endpoint_t(host_id, endpoint_id) ON DELETE CASCADE
);

CREATE INDEX idx_tool_host_endpoint ON tool_t(host_id, endpoint_id);
CREATE INDEX idx_tool_active ON tool_t(active);
CREATE INDEX idx_tool_name ON tool_t(name);

-- Tool Parameters: Defines the arguments for each tool
CREATE TABLE tool_param_t (
    host_id             UUID NOT NULL,
    param_id            UUID NOT NULL,
    tool_id             UUID NOT NULL,
    name                VARCHAR(255) NOT NULL,
    param_type          VARCHAR(50) NOT NULL,      -- 'string', 'number', 'boolean', 'object', 'array'
    required            BOOLEAN DEFAULT true,
    default_value       JSONB,
    description         TEXT,                      -- Helps LLM understand what value to extract
    validation_schema   JSONB,                     -- JSON Schema for complex validation
    order_index         INTEGER DEFAULT 0,
    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, param_id),
    FOREIGN KEY(host_id, tool_id) REFERENCES tool_t(host_id, tool_id) ON DELETE CASCADE
);

-- Skill Dependencies: Manages hierarchies where one skill requires another
CREATE TABLE skill_dependency_t (
    host_id             UUID NOT NULL,
    skill_id            UUID NOT NULL,
    depends_on_skill_id UUID NOT NULL,
    required            BOOLEAN DEFAULT true,
    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY (host_id, skill_id, depends_on_skill_id),
    FOREIGN KEY(host_id, skill_id) REFERENCES skill_t(host_id, skill_id),
    FOREIGN KEY(host_id, depends_on_skill_id) REFERENCES skill_t(host_id, skill_id)
);

-- Agent-Skill Mapping: Links Agents to their Skills
CREATE TABLE agent_skill_t (
    host_id             UUID NOT NULL,
    agent_def_id        UUID NOT NULL,
    skill_id            UUID NOT NULL,

    config              JSONB DEFAULT '{}',
    priority            INTEGER DEFAULT 0,
    sequence_id         INTEGER DEFAULT 0,     -- Order in which skills are concatenated

    aggregate_version    BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, agent_def_id, skill_id),
    FOREIGN KEY(host_id, agent_def_id) REFERENCES agent_definition_t(host_id, agent_def_id) ON DELETE CASCADE,
    FOREIGN KEY(host_id, skill_id) REFERENCES skill_t(host_id, skill_id) ON DELETE CASCADE
);
CREATE INDEX idx_agent_skill_agent ON agent_skill_t(agent_def_id);

-- Skill-Tool Mapping: Implements Progressive Disclosure
CREATE TABLE skill_tool_t (
    host_id             UUID NOT NULL,
    skill_id            UUID NOT NULL,
    tool_id             UUID NOT NULL,

    config              JSONB DEFAULT '{}',
    access_level        VARCHAR(20) DEFAULT 'read', -- e.g., 'read', 'write', 'execute'

    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, skill_id, tool_id),
    FOREIGN KEY(host_id, skill_id) REFERENCES skill_t(host_id, skill_id) ON DELETE CASCADE,
    FOREIGN KEY(host_id, tool_id) REFERENCES tool_t(host_id, tool_id) ON DELETE CASCADE
);
CREATE INDEX idx_skill_tool_skill ON skill_tool_t(skill_id);

-- -- Hindsight Advanced Memory System
-- Transitioned from flat logs to biomimetic memory banks (World, Experiences, Mental Models)

-- Memory bank profiles (Personality & Disposition)
CREATE TABLE agent_memory_bank_t (
    host_id             UUID NOT NULL,
    bank_id             UUID NOT NULL,
    agent_def_id        UUID,                  -- NULL if bank is shared across agents
    user_id             UUID,                  -- NULL if bank is global for the host/agent
    bank_name           VARCHAR(126) NOT NULL,
    disposition         JSONB NOT NULL DEFAULT '{"skepticism": 3, "literalism": 3, "empathy": 3}'::jsonb,
    background          TEXT,
    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, bank_id),
    FOREIGN KEY(host_id) REFERENCES host_t(host_id) ON DELETE CASCADE,
    FOREIGN KEY(host_id, agent_def_id) REFERENCES agent_definition_t(host_id, agent_def_id) ON DELETE CASCADE,
    FOREIGN KEY(user_id) REFERENCES user_t(user_id) ON DELETE CASCADE
);

-- Source documents for memory units
CREATE TABLE agent_memory_doc_t (
    host_id             UUID NOT NULL,
    doc_id              UUID NOT NULL,
    bank_id             UUID NOT NULL,
    original_text       TEXT,
    content_hash        TEXT,
    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY (host_id, bank_id, doc_id),
    FOREIGN KEY (host_id, bank_id) REFERENCES agent_memory_bank_t(host_id, bank_id) ON DELETE CASCADE
);

-- Individual sentence-level memories (The "Atoms" of thought)
CREATE TABLE agent_memory_unit_t (
    host_id             UUID NOT NULL,
    unit_id             UUID NOT NULL,
    bank_id             UUID NOT NULL,
    doc_id              UUID,
    content             TEXT NOT NULL,
    embedding           vector(384),
    context             TEXT,
    event_date          TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
    occurred_start      TIMESTAMP WITH TIME ZONE,
    occurred_end        TIMESTAMP WITH TIME ZONE,
    mentioned_at        TIMESTAMP WITH TIME ZONE,
    fact_type           VARCHAR(32) NOT NULL DEFAULT 'world' CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation', 'mental_model')),
    metadata            JSONB DEFAULT '{}'::jsonb,
    proof_count         INT DEFAULT 1,
    source_memory_ids   UUID[] DEFAULT ARRAY[]::UUID[],
    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, bank_id, unit_id),
    FOREIGN KEY(host_id, bank_id) REFERENCES agent_memory_bank_t(host_id, bank_id) ON DELETE CASCADE,
    FOREIGN KEY(host_id, bank_id, doc_id) REFERENCES agent_memory_doc_t(host_id, bank_id, doc_id) ON DELETE CASCADE
);

CREATE INDEX idx_mem_unit_bank ON agent_memory_unit_t(bank_id);
CREATE INDEX idx_mem_unit_embedding ON agent_memory_unit_t USING hnsw (embedding vector_cosine_ops);

-- Resolved entities (Knowledge Graph Nodes)
CREATE TABLE agent_memory_entity_t (
    host_id             UUID NOT NULL,
    entity_id           UUID NOT NULL,
    bank_id             UUID NOT NULL,
    user_id             UUID,                  -- Link to user_t if this entity is a platform user
    canonical_name      TEXT NOT NULL,
    mention_count       INT DEFAULT 1,
    metadata            JSONB DEFAULT '{}'::jsonb,
    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY (host_id, bank_id, entity_id),
    FOREIGN KEY (host_id, bank_id) REFERENCES agent_memory_bank_t(host_id, bank_id) ON DELETE CASCADE,
    FOREIGN KEY (user_id) REFERENCES user_t(user_id) ON DELETE CASCADE
);

-- Association between memory units and entities
CREATE TABLE agent_memory_unit_entity_t (
    host_id             UUID NOT NULL,
    bank_id             UUID NOT NULL,
    unit_id             UUID NOT NULL,
    entity_id           UUID NOT NULL,
    PRIMARY KEY (host_id, bank_id, unit_id, entity_id),
    FOREIGN KEY (host_id, bank_id, unit_id) REFERENCES agent_memory_unit_t(host_id, bank_id, unit_id) ON DELETE CASCADE,
    FOREIGN KEY (host_id, bank_id, entity_id) REFERENCES agent_memory_entity_t(host_id, bank_id, entity_id) ON DELETE CASCADE
);

-- Cache of entity co-occurrences (Concept Relationship Graph)
CREATE TABLE agent_memory_entity_cooccur_t (
    host_id             UUID NOT NULL,
    bank_id             UUID NOT NULL,
    entity_id_1         UUID NOT NULL,
    entity_id_2         UUID NOT NULL,
    cooccur_count       INT DEFAULT 1,
    last_cooccurred     TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY (host_id, bank_id, entity_id_1, entity_id_2),
    CONSTRAINT entity_cooccur_order_check CHECK (entity_id_1 < entity_id_2),
    FOREIGN KEY (host_id, bank_id, entity_id_1) REFERENCES agent_memory_entity_t(host_id, bank_id, entity_id) ON DELETE CASCADE,
    FOREIGN KEY (host_id, bank_id, entity_id_2) REFERENCES agent_memory_entity_t(host_id, bank_id, entity_id) ON DELETE CASCADE
);

CREATE INDEX idx_mem_cooccur_e1 ON agent_memory_entity_cooccur_t(host_id, entity_id_1);
CREATE INDEX idx_mem_cooccur_e2 ON agent_memory_entity_cooccur_t(host_id, entity_id_2);

-- Links between memory units (Semantic & Causal relationships)
CREATE TABLE agent_memory_link_t (
    host_id             UUID NOT NULL,
    bank_id             UUID NOT NULL,
    from_unit_id        UUID NOT NULL,
    to_unit_id          UUID NOT NULL,
    link_type           VARCHAR(32) NOT NULL,
    weight              FLOAT NOT NULL DEFAULT 1.0,
    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY (host_id, bank_id, from_unit_id, to_unit_id, link_type),
    CONSTRAINT memory_links_type_check CHECK (link_type IN ('temporal', 'semantic', 'entity', 'causes', 'caused_by', 'enables', 'prevents')),
    FOREIGN KEY (host_id, bank_id, from_unit_id) REFERENCES agent_memory_unit_t(host_id, bank_id, unit_id) ON DELETE CASCADE,
    FOREIGN KEY (host_id, bank_id, to_unit_id) REFERENCES agent_memory_unit_t(host_id, bank_id, unit_id) ON DELETE CASCADE
);

-- Directives (Hard rules that override probabilistic learning)
CREATE TABLE agent_memory_directive_t (
    host_id             UUID NOT NULL,
    directive_id        UUID NOT NULL,
    bank_id             UUID NOT NULL,
    name                VARCHAR(256) NOT NULL,
    content             TEXT NOT NULL,
    priority            INT NOT NULL DEFAULT 0,
    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, bank_id, directive_id),
    FOREIGN KEY(host_id, bank_id) REFERENCES agent_memory_bank_t(host_id, bank_id) ON DELETE CASCADE
);

-- Reflections (Synthesized knowledge and high-level observations)
CREATE TABLE agent_memory_reflection_t (
    host_id             UUID NOT NULL,
    reflection_id       UUID NOT NULL,
    bank_id             UUID NOT NULL,
    content             TEXT NOT NULL,
    embedding           vector(384),
    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, bank_id, reflection_id),
    FOREIGN KEY(host_id, bank_id) REFERENCES agent_memory_bank_t(host_id, bank_id) ON DELETE CASCADE
);

CREATE INDEX idx_mem_reflection_embedding ON agent_memory_reflection_t USING hnsw (embedding vector_cosine_ops);

-- Raw Session History (The source of Truth for active conversations)
CREATE TABLE agent_session_history_t (
    host_id             UUID NOT NULL,
    session_id          UUID NOT NULL,
    bank_id             UUID NOT NULL,         -- Links the session to a Hindsight bank
    messages            JSONB NOT NULL DEFAULT '[]'::jsonb,
    metadata            JSONB DEFAULT '{}'::jsonb,
    aggregate_version   BIGINT DEFAULT 1 NOT NULL,
    active              BOOLEAN DEFAULT true,
    update_ts           TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP,
    update_user         VARCHAR(126) DEFAULT SESSION_USER,
    PRIMARY KEY(host_id, bank_id, session_id),
    FOREIGN KEY(host_id, bank_id) REFERENCES agent_memory_bank_t(host_id, bank_id) ON DELETE CASCADE
);

CREATE INDEX idx_session_bank ON agent_session_history_t(host_id, bank_id);