# Why Lians

## The short answer

Lians Guard is a current-state and completion guard for AI coding agents.

> Recover the task. Reject stale state. Block unsupported `done`.

Claude Code, Codex, and other agents can already store instructions and recall
context. Lians is useful when the risk is not forgetting a preference, but
resuming the wrong task state or accepting a completion claim without current
evidence.

## The pressure test

Persistent memory, vector search, MCP, local SQLite, and multi-client support
are useful but not scarce. Native agent products and dedicated memory tools
already cover much of that surface. A broad promise to remember more therefore
creates weak urgency and weak willingness to pay.

The expensive problem is execution state:

- an interrupted session loses the current task and next action;
- a requirement changes while old work still looks current;
- an agent reports success based on its own summary;
- touched files are mistaken for finished work; or
- a reviewer must reconstruct intent, changes, checks, and blockers by hand.

## What Lians is trying to own

The differentiation is a narrow Guard contract:

1. **Recovery.** Restore a bounded current task after interruption or an agent
   switch.
2. **Freshness.** Bind checkpoints to repository and task state, then mark
   mismatches stale.
3. **Evidence.** Separate measured local, measured CI, and human-confirmed proof
   from agent attestations and inferred activity.
4. **Readiness.** Keep the gate closed while criteria are missing or constraints
   are failed, unknown, or blocked.
5. **Human control.** Make the result inspectable and require human review.

## Which approach should I use?

| Need | Best starting point |
|---|---|
| Stable repository instructions | `AGENTS.md`, `CLAUDE.md`, or repository instructions |
| Convenience inside one agent product | That product's native memory |
| A full record of one conversation | Export or retain the transcript |
| A general memory API for an application | A dedicated memory platform |
| Local recovery across supported coding agents | Free Lians recovery |
| Stale-state and evidence-backed review gates | Lians Guard |

Lians is additive. It does not replace source control, CI, the model, repository
instructions, or human review.

## Claim rules

Public comparisons and benchmarks must use identical scenarios, publish the
fixture and configuration, distinguish unsupported from failed, and retire a
claim when a rerun no longer supports it.

Lians can report what its configured policy measured, detected, invalidated, or
blocked. It cannot prove semantic correctness or guarantee safe deployment.

Read [the product direction](product-direction.md), [the Guard contract](lians-guard.md),
and [ContinuityBench v0.1](benchmarks/continuitybench-v0.1.md).