Continuity is core

Many AI tools lose context the moment the UI closes, the tab refreshes, or the operator starts a new thread. tamux is built to avoid that failure mode. The daemon owns the enduring state, so a line of work can survive disconnects and reconnect from another client surface.

Where continuity lives

SQLite state

Threads, tasks, goals, checkpoints, provenance, telemetry, operator model, and collaboration records.

Memory files

SOUL.md, MEMORY.md, and USER.md as curated durable artifacts with strong semantic roles.

Skills

Procedural memory derived from successful workflows and reusable operating patterns.

Recall systems

Search and retrieval over history, telemetry, structural context, and optional external recall integrations.

The persistence split

tamux deliberately uses a persistence split:

  • Structured state: SQLite for threads, tasks, goals, provenance, telemetry, and runtime records.
  • Editable artifacts: markdown memory, skills, ledgers, and selected on-disk outputs.

This keeps the system both durable and operator-auditable. Not everything should be a text file, and not everything should be hidden in a database.

Provenance-backed memory

Durable memory writes in tamux are provenance-backed. The system can track where a fact came from, when it was written, which thread or goal produced it, whether it has been confirmed or retracted, and what later event invalidated it.

Why that matters: memory becomes inspectable and correctable rather than an opaque pile of auto-summaries.
Surface note: the provenance model itself is shipped in the daemon, APIs, and CLI flows today. Richer inspection and confirm/retract experiences exist unevenly across clients, so UI parity should be treated as evolving rather than complete.

Continuity across multi-agent work

Handoffs, visible participants, hidden delegation, subagents, and collaboration sessions all create additional memory-bearing state. Because that state is daemon-owned, multi-agent coordination does not destroy continuity.

  • Handoffs preserve thread context.
  • Goal runners preserve objective-level continuity.
  • Operator profile learning preserves working-style continuity.
  • Provenance preserves trust continuity.

Long-running goals

Goal runners add another layer of continuity by persisting plan state, step state, events, reports, reflections, and checkpoints. That means a long investigation is not just “a big chat” — it is a durable object with inspectable history.

Operator continuity

USER memory is not just a note file. tamux learns structured operator profile state from onboarding, explicit profile edits, behavior patterns, revision history, and interaction signals. That gives future sessions a better model of how the operator works.

What continuity enables

CapabilityEnabled by
Resume work after restartDaemon-owned state + goal/task persistence
Reconnect from another clientShared runtime authority in the daemon
Trust durable factsCurated writes + provenance
Compound skill over timeProcedural memory and generated skills
Preserve multi-agent contextShared memory and thread continuity

Deep storage direction

A more advanced continuity model is that memory should not only remember facts, but know what it knows: source, confidence, lifecycle state, and explicit retraction relationships. tamux is already close to that direction through provenance-backed memory.

What breaks continuity

Continuity gets weaker when operators constantly start fresh threads for the same mission, use chat turns for work that should be a goal run, or treat memory as a dumping ground rather than a curated layer. tamux is designed to preserve continuity, but the workflow still needs to respect the model.