Memoria features

Everything below is implemented in the repository. Quotas and model names come from the README.

Core features

Persistent and cross-session memory
Facts are stored in PostgreSQL with pgvector and retrieved across separate chats.
Personal Intelligence
Toggle global memory access. When off, recall is limited to the current session plus core facts.
MemoryLess incognito sessions
Private chats that never read or write long-term memory. Media slash commands are disabled.
SSE streaming chat
Replies stream through POST /chat/stream, rendered with Markdown and KaTeX.
Forgetting and consolidation
Daily decay and weekly Qwen-Max consolidation.
Conflict detection
Contradicting facts are flagged and superseded during ingestion.
Reflective insights
Periodic Qwen reflection adds patterns about you to the system prompt.
Thumbs up and down feedback
Rate replies to strengthen or weaken the memories that informed them (POST /api/feedback).
Structured output
Qwen JSON responses drive consolidation, reflection and conflict resolution, with schema fallbacks.
Tasks and media library
Create tasks in chat. Browse generated images and videos, download them, or delete them permanently.
Persona
Set response length, tone and behaviour traits. They are added to the system prompt.
Product knowledge base
The AI carries a built-in reference of Memoria's own architecture and features.

Type	Behaviour
`core`	Never decays and is always packed into context first. Manual /memorize facts are stored as core.
`episodic`	Decay rate 0.02.
`semantic`	Decay rate 0.01. Consolidation summaries are stored as semantic memories.
`procedural`	Decay rate 0.005.
`goal`	Decay rate 0.01.
`preference`	Decay rate 0.01.

Type / in chat for this table in the app. In MemoryLess sessions the media commands are disabled.

Command	What it does	Cost
`/`	Show the command help table	Free
`/imagine <prompt>`	Generate an image (wan2.1-t2i-plus)	AI quota
`/gen_video <prompt>`	Generate a video (wan2.1-t2v-turbo)	AI quota
`/gen_voice <prompt>`	Voice overview of the session (Qwen summary + qwen3-tts-flash)	AI quota
`/memorize <fact>`	Store a core memory manually	Free
`/create_task <title>`	Create a pending task	Free
`/tasks_list`	List pending tasks	Free
`/task_complete <ID>`	Mark a task complete	Free
`/list_memory`	List memories (all when Personal Intelligence is on, current chat only when off)	Free
`/forget_memory <ID\|ALL>`	Delete one memory by list ID, or the whole scoped set	Free

Pick a chat model per session from the composer dropdown. The list comes from GET /api/models.

Quotas reset on upgrade. Going over a limit returns HTTP 429. Media uses DashScope default settings, with no per-user size or duration controls.

External agents discover four tools at GET /mcp/memory-skills. Source: backend/app/mcp/memory_skill.py.

Tool	Purpose
`get_core_memories`	Retrieve important memories for a user (importance of 7 or more).
`get_user_preferences`	Extract user preferences from memories.
`forget_memory`	Soft-delete a memory, with a reason.
`strengthen_memory`	Increase the importance of a memory (capped at 10).

Chat, Memories, Persona, Tasks, Media, Settings, Help, Feedback and About.