HINT.

Path-scoped research memory

Leave the next agent a conclusion, not a transcript.

HINT connects what you suspected, what you tried, what you observed, and what future work should do. The record stays readable Markdown; the CLI keeps its links and states honest.

Install HINT See the demo Read the source
agent session
$ hint guide
$ hint src/waiter.ts

## T-request-id
Match responses by request ID.

Based on: N-race
Source: src/waiter.ts.hint

One connected loop

Evidence becomes guidance without losing its route.

  1. H

    Hypothesis

    State a bounded claim, a check, and when the investigation stops.

  2. I

    Iteration

    Take one action that can distinguish the claim from alternatives.

  3. N

    Notice

    Record the result, evidence, and gaps. A saved observation is not proof.

  4. T

    Thesis

    Keep accepted guidance current and supersede it explicitly when conditions change.

Markdown is the source

Readable with or without the CLI.

A file gets a <file>.hint companion. A folder gets _.hint, inherited only by its subtree. Normal reads show accepted Theses; open work and history stay available on demand.

  • Stable IDs and explicit evidence links
  • Atomic writes and revision conflicts
  • No database, daemon, network, or model runtime
---
hint-format: 1
---
# Current theses

## Thesis T-request-id: Match by identity
Status: accepted
Based-on: N-race

### Guidance
Match responses by request ID.

### Rationale
Concurrent requests can share a route.

Optional high-quality mode

Criteria before confidence.

Record the goal, required checks, current working-file fingerprints, known shortcomings, and a subjective score. Passing needs every required criterion, current evidence, no material gap, and a score of at least eight. Ten attempts is a limit, not a success condition.

Node.js 24+

Start with one command.

npm install --global @openhint/cli
hint guide

Then run hint <path> before working and follow the returned NEXT within the user's task and your authority.

What HINT does not pretend to do

HINT does not run checks from Markdown, judge truth, call another model, or guarantee that an agent continues. It validates the record and exposes the next state. The agent still performs the work and reports blockers honestly.