long_term_memory: true without an embedding model crashes the entire server (unrecovered panic in saveCurrentConversation)

knowledgebase.go の Agent.saveCurrentConversation から開始し、agent.go の consumeJob と result.go の JobResult.Finish を通る callback を追跡します。long_term_memory を有効にし、embedding model と knowledge base がない状態で再現し、設定ミスのあるリクエストがエラーを報告しつつ、サーバーを終了させたり他のエージェントに影響を与えたりしないことを確認します。

Hi,

This is found and redacted by Claude, as I'm fucking unable to explain it better.
It's not a code review, but an actual crash I encountered while I let it work for me.

LocalAI version:
localai/localai:v4.9.0-gpu-nvidia-cuda-13 (LocalAI v4.9.0, commit f7ad3f70eb5d8a0ddf80e08557f0d7df28cf032e; vendored github.com/mudler/LocalAGI v0.0.0-20260606071251-14aed1ae4336)

Environment, CPU architecture, OS, and Version:

Linux ia 6.8.0-139-generic #139-Ubuntu SMP PREEMPT_DYNAMIC Sat Aug  1 03:52:05 UTC 2026 x86_64 x86_64 x86_64 GNU/Linux

Bare-metal host (not a VM), Ubuntu 24.04.4 LTS, NVIDIA GeForce RTX 3090, driver 580.173.02, CUDA 13.0. Running via Docker Compose with runtime: nvidia. Backend: llama-cpp / cuda13-llama-cpp.

Describe the bug
Creating an agent via POST /api/agents (or PUT /api/agents/{name}) with "long_term_memory": true but without an embedding model installed/configured causes an unrecovered Go panic (nil pointer dereference) the first time that agent finishes a chat job and the framework tries to save the conversation. The panic is not scoped to the misconfigured agent's own goroutine — it crashes the entire local-ai process (PID 1 in the container). Docker's restart policy (unless-stopped) then restarts the container, which wipes the in-memory observables/job history for every agent on the instance (not just the misconfigured one) and briefly takes the whole API down for all users/agents while it reloads.

Root cause appears to be that long_term_memory depends on an embedding model (default seems to be granite-embedding-107m-multilingual, built-in chromem vector store) that was never installed on this instance, and the save-conversation code path dereferences it without a nil check.

To Reproduce

1. Start LocalAI v4.9.0 with the agent pool enabled (default) and no embedding model installed (fresh instance, or one where granite-embedding-107m-multilingual was never pulled via the gallery).
2. Create an agent:

   curl -X POST http://localhost:8080/api/agents \
     -H "Content-Type: application/json" \
     -d '{
       "name": "test-agent",
       "model": "<any working chat model>",
       "system_prompt": "You are a test agent.",
       "api_url": "http://localhost:8080/v1",
       "api_key": "sk-local",
       "long_term_memory": true
     }'
   

3. Send it a chat message:

   curl -X POST http://localhost:8080/api/agents/test-agent/chat \
     -H "Content-Type: application/json" \
     -d '{"message": "hello"}'
   

4. Wait for the model to generate a response and for the framework to attempt saving the conversation — the whole server crashes and restarts a few seconds later.

Expected behavior
The API should return an error for that specific request/agent (e.g. "long_term_memory requires an embedding model / knowledge base to be configured"), and the rest of the server (other agents, in-flight requests) should be unaffected. At minimum, the panic should be recovered at the job/goroutine level so one misconfigured agent can't take down the whole process.

Logs

Saving conversation agent="supervisor" conversation size=4
panic: runtime error: invalid memory address or nil pointer dereference
[signal SIGSEGV: segmentation violation code=0x1 addr=0x30 pc=0xf9a3a9]

goroutine 7150 [running]:
github.com/mudler/LocalAGI/core/agent.(*Agent).saveCurrentConversation(0x1c1dad1543c0, {0x1c1da66b8f08, 0x4, 0x1c1da591c730?})
	/root/go/pkg/mod/github.com/mudler/[email protected]/core/agent/knowledgebase.go:176 +0x629
github.com/mudler/LocalAGI/core/agent.(*Agent).consumeJob.func8({0x1c1da66b8f08?, 0x4?, 0x4?})
	/root/go/pkg/mod/github.com/mudler/[email protected]/core/agent/agent.go:1420 +0x28
github.com/mudler/LocalAGI/core/types.(*JobResult).Finish(0x1c1da50fd290, {0x0?, 0x0?})
	/root/go/pkg/mod/github.com/mudler/[email protected]/core/types/result.go:43 +0xdf
github.com/mudler/LocalAGI/core/agent.(*Agent).consumeJob(0x1c1dad1543c0, 0x1c1da566e180, {0x48076fd, 0x4})
	/root/go/pkg/mod/github.com/mudler/[email protected]/core/agent/agent.go:1423 +0x2b27
github.com/mudler/LocalAGI/core/agent.(*Agent).run(0x1c1dad1543c0, 0x1c1da4854e70)
	/root/go/pkg/mod/github.com/mudler/[email protected]/core/agent/agent.go:1543 +0xfd
github.com/mudler/LocalAGI/core/agent.(*Agent).Run.func1()
	/root/go/pkg/mod/github.com/mudler/[email protected]/core/agent/agent.go:1521 +0x3a
created by github.com/mudler/LocalAGI/core/agent.(*Agent).Run in goroutine 6913
	/root/go/pkg/mod/github.com/mudler/[email protected]/core/agent/agent.go:1520 +0x1db

docker inspect on the container right after the crash showed ExitCode=0, OOMKilled=false, Error="", and docker events showed a die event with no preceding kill/stop action — confirming this is the process self-terminating on the panic, not an external kill/OOM.

Additional context
Workaround that resolves it: install an embedding model first (POST /models/apply with {"id": "localai@granite-embedding-107m-multilingual"}), then set both long_term_memory: true and enable_kb: true together on the agent (long_term_memory alone is not enough / is what triggers the crash). Retested with the same repro steps after applying the workaround: no panic, container RestartCount unchanged, GET /api/agents/collections shows a collection for the agent, and the job's observables history shows a clean "Recall" step (KB auto-search) before the actual response.

Found while building a multi-agent setup (Planner/Coder/Reviewer/Supervisor) on top of LocalAI's native agent pool.

Regards