Agent Memory Architecture

Three memory types for the LLM agent:

flowchart TB
    subgraph Memory
        Session[Session<br/>goal, progress, steps]
        Stack[Stack<br/>LIFO task queue]
        Facts[Facts<br/>long-term knowledge]
    end

    Session --> SP[System Prompt]
    Stack --> SP
    Facts --> SP
    SP --> Agent[Agent Runner]

Session Memory

Tracks current goal and progress. Cleared on completion.

Field	Description
`goal`	What the user asked for
`progress`	0-100%
`completed`	Actions done
`current_step`	Current work
`next_steps`	Upcoming work
`blockers`	Issues

Stack Memory

LIFO todo queue. Persists across refresh.

Field	Description
`id`	Unique identifier
`description`	Task description
`priority`	high/medium/low
`context`	Optional data

Facts Memory

Long-term knowledge (up to 50 per workspace).

Tools

Session

Tool	Parameters
`set_goal`	`goal`, `steps?`
`update_progress`	`progress`, `completed_action?`
`complete_goal`	`summary`

Stack

Tool	Parameters
`push_task`	`description`, `priority?`, `context?`
`pop_task`	-
`peek_stack`	-
`clear_stack`	-

Facts

Tool	Parameters
`remember`	`message`

Storage Keys

Type	Key
Session	`nodus_agent_session_{workspaceId}`
Stack	`nodus_agent_stack_{workspaceId}`
Facts	`nodus_memories_{workspaceId}`

Files

src/llm/types.ts - Type definitions
src/lib/storage.ts - Storage functions
src/llm/tools/planningTools.ts - Tool registrations
src/canvas/composables/agent/useLLMTools.ts - Tool handlers
src/canvas/composables/agent/systemPrompt.ts - Prompt builder