Cognitive consolidation and trace issues on Linux/Docker setup

[kayne87]

Hi, I found this repo while investigating memory management for LLM agents, and I'm trying this interesting repository, but I'm experiencing a few issues that I hope you can help me resolve 😄

Many thanks in advance

Environment:

• Ubuntu/Debian Docker container
• Python 3.12
• Ollama provider (llama3.2)
• nomic-embed-text
• SuperLocalMemory 3.4.49

Doctor:

root@bab7fa0f01e8:/app# slm doctor
SuperLocalMemory V3 — Doctor (Pre-flight Check)
==================================================

  [PASS] Python: 3.12.13 (>= 3.11)
  [PASS] Core deps: numpy 2.4.4, scipy 1.17.1, networkx 3.6.1, httpx 0.28.1...
  [PASS] Search deps: sentence-transformers, torch, sklearn, geoopt
  [PASS] Dashboard deps: fastapi, uvicorn, websockets
  [PASS] Learning deps: lightgbm 4.6.0
  [PASS] Performance deps: orjson
  [PASS] Embedding worker: responsive (PID 2575, Python /app/.venv/bin/python)
  [PASS] Ollama: running, 2 models, 'llama3.2' available
  [PASS] Disk space: 925.6 GB free
  [PASS] Database: OK (4.12 MB)

Summary: 10 passed, 0 warnings, 0 failed

Status:

root@bab7fa0f01e8:/app# slm status
SuperLocalMemory V3
  Mode: B
  Provider: ollama
  Base dir: /root/.superlocalmemory
  Database: /root/.superlocalmemory/memory.db
  DB size: 4.12 MB

Observed behavior:

1. trace errors in output
2. Cognitive consolidation does not trigger even with many highly-related episodic memories.

Examples:

1. trace errors in output

Trace output seems to have load_active_model failed stable error and sometimes also Query embedding returned None. when it happens all the reported fields are equal to 0

root@bab7fa0f01e8:/app# slm trace "Zebra DataWedge barcode scans Android intent broadcasts"
Query embedding returned None — semantic, hopfield, spreading_activation channels will be skipped this recall
Query embedding returned None — semantic, hopfield, spreading_activation channels will be skipped this recall
Query embedding returned None — semantic, hopfield, spreading_activation channels will be skipped this recall
Query embedding returned None — semantic, hopfield, spreading_activation channels will be skipped this recall
Query embedding returned None — semantic, hopfield, spreading_activation channels will be skipped this recall
load_active_model failed: no such column: bytes_sha256
Query: Zebra DataWedge barcode scans Android intent broadcasts
Type: entity | Time: 119ms
Results: 6

2. Cognitive consolidation does not trigger even with many highly-related episodic memories.

After I have inserted ~6 semantically overlapping memories on the same technical topic

root@c69d6e8d08df:/app# slm consolidate --cognitive
CCQ Cognitive Consolidation
  Clusters processed: 0
  Blocks created:     0
  Facts archived:     0
  Compression ratio:  0.000

I also ran slm reconfigure with a custom profile. The CLI reported:

  + Real consolidation (hnswlib, reversible merges)
  + Inline entity detection (<2 ms trigram lookup)

However, after restart and with multiple related episodic memories, slm consolidate --cognitive still returns the same output.

I am not sure whether I am missing an additional configuration step, or whether the consolidation pipeline is not being triggered as expected.