Architecture

Architecture

Dense-Mem separates memory from the AI assistant.

The assistant still talks to the user and decides how to respond. Dense-Mem owns the memory database, evidence, embeddings, claims, facts, teams, profiles, and API keys.

System Map

flowchart TB
  subgraph Hosts["AI tools"]
    Claude["Claude Code"]
    Codex["Codex"]
    Other["Other MCP clients"]
  end

  subgraph DenseMem["Dense-Mem"]
    MCP["/mcp Streamable HTTP"]
    REST["REST API"]
    UserPortal["User portal /ui"]
    Portal["Control portal"]
    Metrics["Metrics endpoint /metrics"]
    Registry["Tool registry"]
    Memory["Memory orchestration"]
    Recall["Recall service"]
    Embeds["Embedding provider client"]
  end

  subgraph Storage["Storage"]
    Postgres["Postgres teams, profiles, audit"]
    Neo4j["Neo4j graph and vector indexes"]
    Redis["Redis optional shared limits"]
    Prometheus["Prometheus optional telemetry"]
  end

  Hosts --> MCP
  MCP --> Registry
  REST --> Registry
  UserPortal --> REST
  Portal --> REST
  Registry --> Memory
  Registry --> Recall
  Memory --> Embeds
  Recall --> Embeds
  Memory --> Neo4j
  Recall --> Neo4j
  REST --> Postgres
  REST --> Redis
  Metrics --> Prometheus
  REST --> Prometheus
  Portal --> Prometheus

Loading

Responsibility Boundary

Area	Dense-Mem owns	AI assistant owns
Memory writes	Evidence, typed claims, verification, gates, promotion	Deciding what candidate memories to send
Embeddings	Fragment embeddings and recall-query embeddings	No vectors for normal use
Retrieval	Facts, claims, fragments, and clarifications	Choosing how to use recalled context
Corrections	Conflict detection and supersession records	Asking the user which memory is correct
Operations	Teams, profiles, API keys, audit metadata, user portal, control portal	Client-side MCP configuration

Dense-Mem is intentionally narrow. It does not replace the assistant. It gives the assistant a reliable memory service.

Memory Model

Dense-Mem uses three layers:

Layer	Plain meaning	Why it exists
SourceFragment	Original evidence text	Preserves where a memory came from.
Claim	A typed candidate assertion	Lets Dense-Mem check and gate possible memories.
Fact	A promoted active memory	Gives assistants stable memory to recall.

The flow looks like this:

flowchart LR
  Fragment["SourceFragment evidence"] --> Claim["Claim candidate"]
  Claim --> Verify["Verify claim"]
  Verify --> Validated["Validated claim"]
  Validated --> Promote["Promote claim"]
  Promote --> Fact["Active fact"]
  Fact --> Supersede["Supersede after correction"]

Loading

Facts are not overwritten in place. Corrections create new facts and mark older comparable facts as superseded. That preserves history.

Conflict Handling

When a new memory conflicts with an active fact, Dense-Mem should return a clarification instead of guessing.

sequenceDiagram
  participant Assistant
  participant DM as Dense-Mem
  participant User

  Assistant->>DM: remember(candidate memory)
  DM->>DM: verify, gate, compare active facts
  DM-->>Assistant: clarification needed
  Assistant->>User: Which memory should I keep?
  User-->>Assistant: Keep the new one
  Assistant->>DM: confirm_memory(accept_claim)
  DM-->>Assistant: promoted and superseded result

Loading

This is important because memory systems should not silently decide what the user meant.

Team And Profile Isolation

Every memory operation is scoped to the authenticated API key.

An API key belongs to:

• one team
• one named profile
• one profile role, either manager or member
• one scope set, either read-only or read-write

Clients do not need to send a team ID for normal memory routes. Dense-Mem derives the team and profile from the bearer key.

Roles and scopes are separate. Scopes control knowledge access. Manager role controls team metadata and same-team member profile administration. Member keys cannot manage profiles, even with write scope.

Transport

Dense-Mem uses MCP Streamable HTTP at:

/mcp

Supported methods:

Method	Path	Purpose
POST	/mcp	JSON-RPC requests and notifications.
GET	/mcp	Server-to-client SSE stream where supported.

REST and OpenAPI remain available for non-MCP integrations and operations.

Storage Choices

Component	Purpose
Postgres	Teams, profiles, key metadata, audit records, migrations.
Neo4j	Memory graph, claims, facts, relationships, vector indexes.
Redis	Optional for one server; required for multi-instance shared limits.
Prometheus	Optional telemetry store for the /ui app and control portal dashboards.

Embedding Ownership

Dense-Mem embeds normal write and recall text itself.

Clients send text. Dense-Mem sends that text to the configured embedding provider, stores the resulting vectors, and checks that model dimensions stay consistent.

Client-supplied vectors are reserved for advanced search use cases.