Mem0 Architecture - LLM Memory and Context Engineering

Mem0 is an incremental memory pipeline that runs in real time alongside conversations. The overall flow consists of two stages: Extraction and Update.

Inputs and Outputs

Each new message pair (m_{t-1}, m_t) triggers one execution:

Input: New message pair (m_{t-1}, m_t) + conversation summary S + last m messages
Output: Updated memory store M'

Extraction Stage

The extraction stage constructs full context and calls the LLM to extract candidate facts.

Component	Role
Conversation summary S	Global topic understanding, refreshed asynchronously
Last m messages	Fine-grained recent temporal context
Current message pair	New interaction to process

The extraction function φ(P) outputs a set of candidate facts Ω = {ω1, ω2, ..., ωn}.

Asynchronous summary: Conversation summaries are refreshed periodically by an independent module without blocking the main pipeline, ensuring extraction always has up-to-date semantic context.

Update Stage

For each candidate fact ω_i, the system first retrieves the top-s most similar existing memories via vector search, then decides one of four operations through LLM Tool Call.

Operation	Trigger Condition	Effect
ADD	No semantically similar memory exists	Create new memory entry
UPDATE	Existing memory needs supplementation	Merge into existing entry
DELETE	Contradicts existing memory	Mark or remove old entry
NOOP	Already exists or insignificant	No modification

Update Algorithm (Pseudocode)

for each fact f in F:
  operation = ClassifyOperation(f, M)
  if operation == ADD:
    M = M ∪ {(new_id, f, "ADD")}
  elif operation == UPDATE:
    m_i = FindRelatedMemory(f, M)
    if InformationContent(f) > InformationContent(m_i):
      M = (M \ {m_i}) ∪ {(id_i, f, "UPDATE")}
  elif operation == DELETE:
    m_i = FindContradictedMemory(f, M)
    M = M \ {m_i}
  # NOOP: no action
return M

Experimental Parameters

Parameter	Value	Description
Recent message window m	10	Number of historical messages participating in extraction
Similar memory count s	10	Number of candidates for conflict detection
Extraction / update model	GPT-4o-mini	Temperature set to 0
Vector index	Dense vectors + similarity search	Supports efficient recall

Why Not Use a Classifier

Mem0 does not train an additional operation classifier; instead, it lets the LLM reason directly and choose appropriate operations between candidate facts and similar memories. Reasons:

Semantic relationships are highly context-dependent; rules or lightweight classifiers cannot cover all cases
Structured decisions output via tool calling are easy to integrate
Reduces additional training and maintenance costs

Basic memory suffices for single-hop and multi-hop QA; for cross-entity relational reasoning, see Mem0g Graph Memory.

​Inputs and Outputs

​Extraction Stage

​Update Stage

​Update Algorithm (Pseudocode)

​Experimental Parameters

​Why Not Use a Classifier

Inputs and Outputs

Extraction Stage

Update Stage

Update Algorithm (Pseudocode)

Experimental Parameters

Why Not Use a Classifier