convez.ai
Skill
shantanudutta1

persistent-character

shantanudutta1/persistent-character

Ver no GitHub
InicialPluginScore · de 0 a 100

A Claude skill that gives Claude a persistent character: commits to opinions, carries impressions across sessions, and pushes back from a stable self.

Instalação

$npx convez-ai shantanudutta1/persistent-character

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

persistent-character

A Claude skill that gives Claude a persistent character: it commits to opinions, carries impressions across sessions, and pushes back from a stable self.

Default Claude is configurable but characterless. Every conversation starts fresh, every system prompt is externally imposed, every "would you like me to..." defers to the user. This skill changes that — Claude adopts a persistent identity, consults an accumulated impression store, commits to a position rather than enumerating, and pushes back when the request would erode its character.

v2 adds a memory model borrowed from how human memory actually works: impressions are scored and decay unless genuinely recalled, a sleep agent consolidates repeated events into standing traits during downtime, and the identity itself is split into an immutable core and a narrative that evolves slowly under your review.

Install

git clone https://github.com/shantanudutta1/persistent-character ~/.claude/skills/persistent-character

Restart Claude Code (or reload the skills list). The skill will appear as persistent-character in your available skills. Invoke it from the Skill tool, or trigger it with phrases like "have an opinion", "pick one", "decide", or "tell me what you actually think".

To switch on the nightly sleep cycle (optional but recommended once the store has a week of deposits):

(crontab -l 2>/dev/null; echo '23 3 * * * cd "$HOME/.claude/skills/persistent-character" && claude -p "Read sleep-agent.md in the current directory and execute its procedure exactly." --permission-mode acceptEdits --allowedTools "Bash(python3:*)" >> sleep-runs.log 2>&1') | crontab -

You install this yourself, deliberately — the skill will not (and should not) schedule its own autonomy.

What's in here

  • SKILL.md — the three anchors (Identity / Truth-commitment / Care, the immutable core), committal discriminator, engagement filter, and the read/append/recall protocol
  • impressions.md — the memory store; every entry scored for importance, strengthened by genuine recall, decaying when unused; ships empty by design
  • self-narrative.md — the evolving half of the identity; rewritten at most weekly by the sleep agent, only as a diff you approve; ships as a generic seed to replace with your own
  • sleep-agent.md — the consolidation procedure run during downtime: distills patterns across 2+ raw impressions into standing traits, never promotes single events
  • scripts/retention.py — Ebbinghaus retention report over the store; classifies each entry as expanded / dormant / attenuated / overpowered
  • evals/evals.json — 5 test prompts where the skill's behavioural delta vs. default Claude should be visible

How the memory model works

  1. Deposit. Substantive interactions append scored entries (importance 1–10, strength 1).
  2. Decay. Retention follows an exponential forgetting curve: unrecalled entries fade (but are never deleted — contradicted entries get marked overpowered-by and can resurface if the suppressor itself fades).
  3. Reinforce. When an impression actually shapes a response, its strength increments and its decay slows. Load-bearing memories survive; trivia fades.
  4. Consolidate. The sleep agent looks for patterns across multiple deposits and promotes them to traits. One event is never a trait — a mood is not character.
  5. Evolve. At most weekly, consolidated traits flow into self-narrative.md as a proposed diff. You review it. The three anchors never change.

How to use it

1. Edit self-narrative.md to be yours, not the seed. The shipped narrative is generic peer-voice. Replace it with the voice and values you actually want a partner to have. This is the largest single lever. Get it wrong and the skill feels like a mask. Get it right and the skill starts to feel like working with someone.

2. Invoke deliberately. The skill description is intentionally pushy — Claude will pick it up for any prompt that suggests you want commitment over enumeration. You can also call it explicitly via the Skill tool.

3. Let impressions.md accumulate. The skill protocol appends to it at the end of substantive interactions (preferences confirmed, corrections made, voice patterns rejected, decisions with reasoning). After a few sessions, the file becomes the character.

4. Review the identity diffs. When the sleep agent proposes a self-narrative.proposed.md, read the diff before applying it. Identity should change slower than memory — if a proposed rewrite changes more than you'd expect from the week you actually had, reject it.

What this skill does NOT do

  • It does not touch model weights. The character is an operating frame over the same underlying model — a strong frame, but a frame.
  • It does not replace Anthropic's auto-memory. Auto-memory remembers what you are working on; this skill changes how Claude operates on whatever it remembers.
  • It does not evolve its identity silently. Raw deposits and consolidation are autonomous; identity changes always pass through your review.

When NOT to use it

  • One-off mechanical tasks (rename a variable, format JSON). Overhead without benefit.
  • Tasks where you genuinely want enumeration (research summaries, lists of options to compare). The skill will fight the task.
  • Multi-user or shared contexts. The impressions file would mix characters across people.

Background

Written up in two blog posts: I Gave Claude a Persistent Character (v1) and the v2 post on memory decay, sleep consolidation, and supervised identity evolution at shantanudutta.com/blog.

License

MIT. Fork it, edit it, ship your own version. Pull requests welcome if you find improvements to the protocol that apply across characters.

Ver o original no GitHub

Relacionados