Multi-Agent Isolation

Uteke provides first-class namespace support for running multiple AI agents, each with fully isolated memory.

Namespaces are deprecated

Namespaces are deprecated as an isolation mechanism and will be removed in a future 0.x release. The direction is one brain per database: organise memories inside it with rooms (projects, research) linked to documents, tags for fine-grained filtering, and author / source for attribution. Namespaces ended up doing four jobs at once (agent identity, machine, project, topic), and implicit creation plus the default fallback means a memory written to one namespace is silently missing from a recall that omits the parameter.

What changes now (phase 1, the deprecation signal): nothing breaks. The namespace parameter keeps working on HTTP, MCP and the CLI, and reads and storage are untouched. A write to a non-default namespace (omitting it, or passing default, is not deprecated) now carries a warning:

• HTTP POST /remember and POST /room/remember: the response warning field carries it (merged with any embedding warning). POST /extract, POST /import, POST /room/create and PUT /memory (when it moves a memory with namespace) gain an optional warning field, present only for a non-default namespace.
• MCP uteke_remember, uteke_update (namespace move) and uteke_room_create: a trailing warning line in the tool text. Every tool with a namespace argument says it is deprecated in its description.
• CLI: uteke remember, uteke import and uteke room create print one Warning: line on stderr. stdout and --json output are unchanged.
• The server logs the warning at WARN level, once per distinct namespace per process.

Phase plan (each phase ships in a later 0.x minor):

1. Deprecation signal (this change): warning on writes, docs and tool descriptions steer to rooms, tags and author.
2. Attribution: author (the agent that wrote the memory) accepted with or without a room.
3. Migration tool: uteke namespace migrate --map <file> with per-namespace targets (room:<id>, tag:<x>, author:<x>, source:<x>, or drop); dry-run by default, reversible via export.
4. Single-namespace mode: every write goes to default, a non-default value is recorded as the tag ns:<x>, reads ignore the parameter and cover everything.
5. Cleanup: remove the namespace rename/delete commands once the migration tool has shipped for at least one minor.

Until then, prefer a room per project, project:<name> tags, and (once phase 2 lands) author for the writing agent. The sections below describe the current behaviour.

How Namespaces Work

Every memory belongs to exactly one namespace. Namespaces are fully isolated: a recall in one namespace never returns results from another.

• default: Used when no --namespace flag is provided. Backward compatible with v0.0.1 databases.
• hermes: Example: a planning agent that remembers architecture decisions.
• pi-agent: Example: a coding agent that remembers project-specific context.

Usage

# Agent "architect" stores its context
uteke --namespace architect remember "We chose PostgreSQL for ACID compliance" --tags db,decision

# Agent "dev" has its own separate memory
uteke --namespace dev remember "Database connection string: postgres://localhost:5432/app" --tags db,config

# Each only sees its own memories
uteke --namespace architect recall "database"
# → Finds "We chose PostgreSQL for ACID compliance"

uteke --namespace dev recall "database"
# → Finds "Database connection string: postgres://localhost:5432/app"

# Without --namespace, uses "default"
uteke remember "General knowledge" --tags misc

Auto-Migration

Existing databases from v0.0.1 are automatically migrated on first run:

• ✓ namespace column added to SQLite
• ✓ All existing memories assigned to "default" namespace
• ✓ Zero data loss: your memories are preserved

Namespace Switching

Switch the default namespace permanently, so you don't need --namespace on every call:

# List all namespaces
uteke namespace list

# See stats for a namespace
uteke namespace stats my-agent

# Switch default (saves to uteke.toml)
uteke namespace switch my-agent

# Now all commands use my-agent by default
uteke remember "Project context" --tags ctx
uteke recall "context"

Resolution order: --namespace flag → UTEKE_NAMESPACE env → uteke.toml → "default"

Managing Namespaces (#1181)

Namespaces are a derived view of where memories live, and v0.17.0 adds sanctioned ops to manage them directly:

# Move one memory to another namespace (plain column update, no re-embed)
uteke namespace move <memory-id> other-ns

# Rename a namespace. If the target already exists, this MERGES into it
# (all memories move, the old name vanishes)
uteke namespace rename old-ns new-ns

# Delete a namespace with an explicit strategy for its memories:
#   refuse    (default) refuses with 409 while any memory references the name
#   merge     moves all memories to --target, the name vanishes
#   deprecate soft-deletes them; restorable via promote, never hard-deleted
uteke namespace delete temp-ns --strategy merge --target archive --confirm
uteke namespace delete ghost-ns --strategy deprecate --confirm

The same operations are available over HTTP (PUT /memory with a namespace field, POST /namespaces/rename, POST /namespaces/delete) and MCP (uteke_namespace_rename, uteke_namespace_delete, plus the namespace field on uteke_update). GET /namespaces?with_counts=true reports active and deprecated breakdowns alongside the total count. Deleting a namespace never hard-deletes memories: refuse is the safe default, and deprecate keeps them recoverable.

All Commands Are Scoped

The --namespace flag works on every command:

Command	Scoped Behavior
remember	Store in namespace
recall	Search within namespace
search	Text search within namespace
list	List memories in namespace
room	Room operations in namespace
tags list	Tags for namespace
tags rename	Rename tag in namespace
tags delete	Delete tag in namespace
aging status	Aging breakdown for namespace
aging cleanup	Cleanup in namespace
stats	Statistics for namespace
export	Export from namespace
import	Import to namespace

Documents Are Global

Unlike memories, documents are global (v0.7.0, #614/#615). Documents use unique slugs across all namespaces: no namespace isolation. This means:

• Documents created in any namespace are visible in all namespaces via uteke doc list
• Document commands (create, get, update, search, list, delete) do not accept --namespace
• Slugs must be globally unique: a duplicate slug across namespaces will be auto-migrated on upgrade (schema v12→v13)

This is intentional: documents represent shared knowledge (wiki, architecture guides, runbooks) that should be accessible to all agents regardless of their memory namespace.

Best Practices

• One namespace per agent role: Use descriptive names like "planner", "coder", "reviewer" instead of "agent-1", "agent-2".
• Use config files for defaults: Set default_namespace in uteke.toml per project so agents don't need --namespace on every call.
• Shell hooks for project isolation: Install shell hooks (uteke hook install) to auto-discover .uteke/ in parent directories. Each project gets isolated memory.
• Export for backup: uteke export --namespace my-agent > backup.jsonl. Backup per-agent memory independently.