Project state survives
Progress, decisions and relevant memory persist independently from any single model session or worker.
Tenrec by Veritiana
Persistent AI software runtime. The work persists. Workers do not.
Tenrec is a persistent AI software runtime built around short-lived specialized workers and selective project memory. The project keeps its validated state while each worker receives only the context needed for bounded work.
Persistent
Project state survives individual workers.
Selective
Only relevant memory is routed into work.
Governed
Accept, retry, escalate or stop explicitly.
Runtime model
State persists. Workers do not.
Persistent project state
Progress · decisions · selective memory · validated skills
Intent router
Recognize the task
Maps incoming work to a known task archetype.
Planner
Build the route
Selects workers, tools, checks and execution order.
Memory router
Load only what matters
Retrieves compact relevant memory instead of full history.
Short-lived specialist workers
Explorer · Implementer · Reviewer · Test runner · Domain worker
Deterministic evaluation
Tests and validators decide whether work passed.
Governance
Accept, retry, escalate or stop based on explicit policy.
Execution traces
Every worker action and outcome remains observable.
How Tenrec works
Follow the persistent runtime step by step.
This walkthrough explains how a software goal becomes bounded specialist work, how failures become explicit repair tasks, and how validated outcomes persist as project memory.
Architecture walkthrough
Gmail → classify → Slack
User goal
Connect Gmail and Slack.
Detect important incoming emails, classify them and post relevant alerts to Slack.
Keep the integration reliable and testable.
Tenrec understands the work
The request becomes an execution architecture.
Classified work
External service integration
Connector · processing · routing · delivery
Repository inspection
Existing project examined
Services · jobs · config · tests · conventions
Work graph
Specialist tasks created
Inspect → build → classify → deliver → verify
Success contract
Validation defined
Connector tests · delivery tests · project constraints
Short-lived workers
Bounded specialist work.
Persistent state
The project survives.
Validation first
Evidence decides success.
Governed execution
Explicit next action.
Validated learning
Useful results compound.
Why Tenrec exists
Long-running software work needs persistence without a long-running worker.
Workers stay replaceable
Specialists are created for bounded work, then disappear. No single long-lived worker becomes the project.
Memory is routed
Tenrec does not keep injecting an ever-growing transcript. It selects compact memory relevant to the current task.
Software decides success
Deterministic validators and governance rules decide whether an execution is accepted, retried, escalated or stopped.
Architecture
A bounded loop instead of an autonomous black box.
Tenrec decomposes AI software work into explicit layers so planning, memory, execution, evaluation and governance can be inspected independently.
01 · Intent
Recognize the work
A compact classifier maps incoming work to known task archetypes and gives the planner an explicit starting point.
02 · Plan
Choose a bounded route
The planner selects worker roles, tools, required context, validators and a concrete execution sequence.
03 · Memory
Load relevant context
The Memory Router chooses compact topic memory instead of flooding every worker with the whole project history.
04 · Execute
Run specialist workers
Short-lived workers perform bounded tasks with only the tools and context required for that role.
05 · Evaluate
Test the result
Tests, validators and structured checks decide whether the result satisfies the actual task contract.
06 · Govern
Control the next action
Governance policy chooses accept, retry, escalate or stop instead of allowing an unbounded repair loop.
07 · Observe
Trace every execution
Runtime traces expose worker activity, decisions, costs, validations, retries and governance outcomes.
08 · Improve
Learn under control
Useful memory and skills can be retained selectively while obsolete context is rotated instead of accumulating forever.
Selective memory
Memory should be useful, not merely large.
Tenrec separates project persistence from prompt context. A small Memory Router decides which compact memory blocks are relevant to the current task.
Topic-oriented compact memory
Learned routing instead of loading everything
Rotation and refresh instead of endless accumulation
Persistent project state remains independent from model context
Current task
Fix failing API integration tests
Memory Router
Select only relevant project memory
Loaded
API contract
Loaded
Recent test failures
Not loaded
Unrelated UI history
Governance
No endless repair loop.
Evaluation produces evidence. Governance decides what happens next. The runtime does not let a worker simply keep trying until tokens run out.
Accept
Result satisfies the contract.
Retry
A bounded correction is justified.
Escalate
Human or higher-level decision is required.
Stop
Further autonomous work is not justified.
Observability
See exactly what the runtime did.
Tenrec exposes execution traces, quality signals and governance outcomes so AI software work can be debugged like software.
Worker activity
who did what
Evaluation
what passed or failed
Governance
why execution continued or stopped
Cost and traces
what the loop consumed
Interfaces
Operate it from the CLI. Watch it live in the UI.
CLI
Developer-first control
$ tenrec run "implement API validation"
intent → plan → memory → workers → evaluate → govern
Live UI
Activity, quality and cost
Live
Worker messages
Trace
Execution path
Quality
Evaluation signals
Open intelligence layer
The routing and planning components are inspectable.
Tenrec separates learned recognition from runtime control so each layer can be tested independently.
Open-source AI software runtime
Build persistent AI software work without a persistent agent.
Tenrec keeps project state persistent while short-lived specialized workers operate with selective memory, deterministic evaluation, governance and full execution visibility.
Veritiana