Memory layer across agents

What broke — and how it was caught

Every one of these failed silently. The tools returned success while storing nothing, which is the worst failure mode a memory layer can have: you only find out weeks later, when the thing you told it is gone.

1. DEFECT 01

   Reasoning models returned empty content

   SYMPTOM

   Tools reported success while storing nothing. Zero facts extracted, no error anywhere.

   CAUSE

   mem0 sends num_predict=2000. Reasoning models spend the whole budget in their thinking channel, which Ollama strips from message.content — the response arrives done_reason="length" with an empty string. The upstream fix appends /no_think to the prompt text, which stopped working on Ollama 0.21+.

   FIX

   Set think=False at the API level, and prefer a non-thinking instruct model.

   EVIDENCE

   Against mem0's real fact-retrieval prompt, 3 inputs: qwen2.5:3b-instruct returned valid JSON 3/3, qwen3:4b 1/3.

2. DEFECT 02

   The default prompt discarded every technical fact

   SYMPTOM

   A memory layer built for engineering context stored coffee preferences and dropped engineering statements.

   CAUSE

   mem0's FACT_RETRIEVAL_PROMPT opens "You are a Personal Information Organizer". Measured: "I prefer dark roast coffee, I work as a software engineer" extracted two facts; "The batch scheduler fix raised successfully placed jobs from 232 to 247" extracted an empty list.

   FIX

   A replacement extraction prompt written for engineering statements, wired in through custom_fact_extraction_prompt.

   EVIDENCE

   Silent total failure for the layer's actual purpose — success responses, empty store.

3. DEFECT 03

   delete_all_memories ignored its own guard

   SYMPTOM

   A no-argument call wiped the entire default user scope, despite a docstring promising at least one filter is required.

   CAUSE

   The scope check ran against a value that had already been defaulted, so the guard could never be true — dead code. The sibling delete_entities checks the raw parameters and is correct; the two disagreeing is what makes it a bug rather than a design choice.

   FIX

   Wrap the registered tool and restore the documented behaviour, with an explicit environment-variable opt-out.

   EVIDENCE

   Test discriminates: guard on with no args refuses and 2 memories survive; guard off with no args deletes 2, reproducing the bug; guard on with explicit scope deletes 6, so the intended path still works.