Summary

memu-py 0.1.8 on PyPI still ships the old memu.server + RecallAgent architecture that was removed in the v1.5.0 refactor (#81, commit 4a2e86c). Users who pip install memu-py and run python -m memu.server.cli start get a server that cannot retrieve any memories.

Related: #60, #8

Root Cause

The old RecallAgent.retrieve_relevant_memories() reads from a file-system memory_dir ({memory_dir}/{agent_id}/{user_id}/*_embeddings.json), but MemoryAgent.run() writes to PostgreSQL (memory_items table with pgvector embeddings). The read and write paths are completely disconnected.

Additionally, cli.py hard-exits if OPENAI_API_KEY is not set, even though the embedding data is already stored in PG and does not require OpenAI for retrieval.

Diagnosis

# PG has data:
SELECT count(*) FROM memory_items;  -- 326 rows, all with embedding vectors

# But API returns nothing:
POST /api/v1/memory/retrieve/related-memory-items
{"query": "test", "user_id": "boris", "top_k": 5}
→ {"related_memories": [], "total_found": 0}

Why: RecallAgent initializes MemoryFileManager pointing at a file directory that is empty. It never touches PG.

Our Workaround

We patched memory_service.py to bypass RecallAgent entirely and query PG pgvector directly:

# In retrieve_related_memory_items():
# 1. Embed query with local bge-m3 (1024-dim, matching stored vectors)
# 2. Query PG directly:
SELECT id, summary, memory_type, created_at,
       1 - (embedding <=> $query_vec::vector) AS similarity
FROM memory_items
WHERE user_id = $user_id
  AND 1 - (embedding <=> $query_vec::vector) >= $min_similarity
ORDER BY embedding <=> $query_vec::vector
LIMIT $top_k

This returns correct results immediately.

Suggestion

1. Short term: Publish a new PyPI release that either (a) removes the deprecated memu.server package, or (b) updates it to use the new src/memu/app/retrieve.py workflow
2. Long term: The v1.5.x architecture looks correct — just needs a PyPI release. Note: v1.5.x requires Python ≥3.13, which blocks users on 3.11/3.12

Environment

• memu-py==0.1.8 from PyPI
• Python 3.11.9, Windows 11
• PostgreSQL 18.3 + pgvector + TimescaleDB
• Embedding: local bge-m3 (1024-dim) via sentence-transformers