Memoria, a personal AI that remembers, forgets and reflects

What it does

• Persistent memory

  Facts from conversations are embedded with text-embedding-v3 (1024 dimensions) and stored in PostgreSQL with pgvector. Six memory types: core, episodic, semantic, procedural, goal, preference.

• Cross-session recall

  Retrieval ranks memories by vector similarity × importance × recency, so the agent can use what you said in earlier, separate chats.

• Personal Intelligence and MemoryLess

  Turn global memory access on or off. MemoryLess incognito chats never read or write long-term memory.

• Forgetting and consolidation

  Daily exponential decay archives stale memories. A weekly Qwen-Max job merges related facts into short summaries.

• Conflict detection

  A new fact that contradicts a stored one is flagged and supersedes it during ingestion.

• Reflection

  Qwen periodically reflects on stored memories and adds higher-level patterns about you to the system prompt.

• Slash commands and streaming chat

  Type / for nine commands. Replies stream token by token over SSE with Markdown and KaTeX.

• MCP memory skills

  Four memory tools are published at GET /mcp/memory-skills so external agents can read and curate memory.

See every feature

How the pieces fit

1. 1. Chat

   FastAPI loads the Redis session, retrieves memories from pgvector, and asks Qwen for a reply.

2. 2. Ingest

   A Celery worker extracts facts, embeds them and stores them with conflict checks.

3. 3. Maintain

   Celery Beat runs daily decay and weekly consolidation.

4. 4. Reflect

   Reflection runs in the background on every 10th user message.

Read the memory lifecycle

Benchmark result

Two scenarios scored slightly lower with memory (-3.7%) and three were unchanged. See method and all 12 results.

Screenshots