Skip to main content

memory

Cognitive memory system that gives agents persistent awareness. Agents track goals, plans, tasks, notes, and facts — all surviving context compaction. The agent always knows what it's doing and what remains.

PropertyValue
Module IDmemory
Version1.0.0
Typesystem (auto-injected into system prompt)
DependenciesNone

Design Philosophy​

  • Always visible — working memory is injected into the system prompt at every turn, including after compaction. The agent never loses its cognitive state.
  • Opt-in layers — each memory layer (working, todo, episodic, semantic, procedural) is independently activatable. Zero overhead if not enabled.
  • Agent-driven — the system provides tools, not commands. The agent decides when to set goals, create tasks, or store facts.
  • Adaptive — behavior instructions adapt to task complexity. Simple questions get direct answers. Complex tasks get full workflow guidance.

Configuration​

modules:
  memory:
    config:
      working_memory: true
      todo_list: true
      checkpoint: true
      episodic: true
      semantic:
        vector: true
        graph: true
      procedural: true
      runtime:
        goal_guardian: true
        content_cache: true
      limits:
        max_injected: 7
        budget_percent: 15

Actions (16)​

set_goal​

Set the main goal and optional sub-goals. Risk: low

set_plan​

Set the step-by-step plan. Automatically creates todo items for each step. Risk: low

update_plan_step​

Advance to the next step or set a specific step index. Risk: low

add_todo​

Add a task to the todo list. Risk: low

update_todo​

Update task status: pending, in_progress, done, blocked. Returns the full task list with progress. Risk: low

note​

Add a sticky note — something the agent must not forget before finishing. Risk: low

resolve_note​

Mark a sticky note as done. Risk: low

checkpoint​

Save a progress checkpoint — a self-assessment summary of current state. Risk: low

add_fact​

Store a key fact that survives compaction. Risk: low

track_entity​

Track an active entity (file, person, concept) in working memory. Risk: low

add_relationship​

Add a relationship between two entities in the knowledge graph. Risk: low

recall​

Search memory for relevant facts and context. Risk: low

forget​

Remove a specific fact by ID. Risk: low

cache_content​

Cache file content for instant O(1) access on subsequent reads. Risk: low

get_snapshot​

Get the full memory snapshot across all active layers. Risk: low

add_episode​

Record a session episode in episodic memory. Risk: low