Known Issues & Patches

mem0 has known upstream bugs that affect AWS Bedrock + OpenSearch / S3 Vectors usage. PRs have been submitted but are not yet merged. You must apply patches manually before using this service.

Summary

Issue	PR	Affects	Status
OpenSearch 3.x nmslib engine deprecated	#4392	OpenSearch 3.0+	Pending merge
Converse API temperature + top_p conflict	#4393	Claude Haiku 4.5 and newer models	✅ Merged via #4469
S3Vectors invalid filter format	#4554	S3 Vectors backend	Pending merge
MiniMax models not recognized as valid provider	#4609	All MiniMax models on Bedrock	✅ Merged 2026-03-30
Telemetry causes thread leak	— (config, no PR needed)	All deployments	✅ Fix: set MEM0_TELEMETRY=false

Configuration Issue: Telemetry Thread Leak

mem0 enables anonymous telemetry by default. Every add / search / delete call triggers capture_event() in mem0/memory/telemetry.py, which:

1. Creates a new AnonymousTelemetry instance → a new Posthog client
2. Each Posthog client spawns a Consumer background thread
3. The thread is a daemon but never exits — it blocks in queue.get() indefinitely
4. Data is uploaded to https://us.i.posthog.com (mem0 official PostHog), including your collection name, LLM type, and vector store type

Impact: After a long auto_dream run (hundreds of add/delete calls), we observed 135 zombie PostHog threads. Thread count grows ~1/request and will eventually destabilize the process.

Fix — add to .env:

MEM0_TELEMETRY=false

Then restart the service:

sudo systemctl restart mem0-memory.service

Thread count drops from 135 → 8 immediately after restart.

TIP

This is an official config option supported by mem0. Private/self-hosted deployments have no reason to send telemetry to mem0's servers.

PR #4392: OpenSearch 3.x nmslib Engine Deprecated

Patch steps:

# Locate the file
python3 -c "import mem0; import os; print(os.path.join(os.path.dirname(mem0.__file__), 'vector_stores/opensearch.py'))"

# Replace nmslib → lucene
sed -i 's/"engine": "nmslib"/"engine": "lucene"/g' <path>

PR #4393: Converse API temperature + top_p Conflict

✅ Resolved: Fixed in upstream via PR #4469 (merged 2025-03-25). Run pip install --upgrade mem0ai — no manual patch needed.

Claude Haiku 4.5 and newer models reject requests that include both temperature and top_p simultaneously. mem0 defaults top_p=0.9, causing ValidationException on Bedrock Converse API calls.

Patch steps:

# Locate the file
python3 -c "import mem0; import os; print(os.path.join(os.path.dirname(mem0.__file__), 'llms/aws_bedrock.py'))"

Edit the file: comment out the topP line in the Converse API inferenceConfig block. Also change top_p default to None in mem0/configs/llms/aws_bedrock.py.

PR #4554: S3Vectors Filter Format

s3_vectors.py's _convert_filters() generates an incorrect filter format for the S3Vectors query_vectors API. It produces {"equals": {"key": "...", "value": {"stringValue": "..."}}} instead of the required MongoDB-style {"field": {"$eq": "value"}}.

Patch steps:

# One-click patch (provided by this project)
python3 patch_s3vectors_filter.py

PR #4609: MiniMax Models Not Recognized on AWS Bedrock

✅ Resolved: Fixed in upstream via PR #4609 (merged 2026-03-30). Run pip install --upgrade mem0ai — no manual patch needed.

mem0's aws_bedrock LLM provider had three bugs affecting MiniMax M2.x models on Bedrock Converse API:

Bug 1 — PROVIDERS allowlist
minimax was missing from the list, causing ValueError: Unknown provider in model at startup.

Bug 2 — Reasoning model response format
MiniMax M2.5 (and M2.1) are reasoning models. Their Converse API response includes a reasoningContent block before the actual text block. Taking content[0]["text"] directly raised a KeyError.

Bug 3 — System messages discarded
The original code only forwarded the last user message to the Converse API, silently dropping role=system messages. Without the JSON instruction, MiniMax returned free-form markdown, causing json.JSONDecodeError in mem0's fact-extraction pipeline.

Why MiniMax M2.5?

We benchmarked MiniMax M2.5 against Claude Haiku 4.5 and DeepSeek V3.2 for mem0's fact-extraction workload (short text in, short JSON out) on AWS Bedrock us-east-1:

Model	Avg latency	Input price	Output price	Notes
Claude Haiku 4.5	~1.0s	$1.00 / 1M tokens	$5.00 / 1M tokens	Fastest; most expensive
DeepSeek V3.2	~2.4s	$0.62 / 1M tokens	$1.85 / 1M tokens	No clear advantage over MiniMax
MiniMax M2.5 ✅	~2.6s	$0.30 / 1M tokens	$1.20 / 1M tokens	Best cost; ~3× cheaper than Haiku

MiniMax M2.5 is our default choice: ~3× cheaper than Claude Haiku 4.5 at an acceptable ~2-3 s latency for background memory extraction tasks.

TIP

If raw speed matters more than cost, switch back to Claude Haiku 4.5 by setting LLM_MODEL=us.anthropic.claude-haiku-4-5-20251001-v1:0. No patch needed for Haiku.

WARNING

minimax.minimax-m2 (non-reasoning variant) does not work — its Converse API response contains only reasoningContent with no text block. Use M2.1 or M2.5 instead.

After PRs Are Merged

Once all PRs are merged upstream, simply upgrade mem0 and patches are no longer needed:

pip install --upgrade mem0ai

Check PR status:

gh pr view 4392 --repo mem0ai/mem0 --json state -q .state
gh pr view 4554 --repo mem0ai/mem0 --json state -q .state
# PR #4609 (MiniMax) — merged 2026-03-30, no longer needs tracking

WARNING

After every pip install --upgrade mem0ai, re-apply patches until the corresponding PR is merged.