agent-inner-life

Your agent remembers how things have been going.

An agent skill that gives an agent dated state, an evening journal, free thinking at night, and a way to read all of it back.

  something happens   →   state.md           dated facts, no scores
                               ↓
  evening             →   journal entry  +   a short summary in memory
                               ↓                        ↓
  weekly              →   rollup             tomorrow's system prompt
                               ↑                        ↓
  night               →   a dream        →    back into state

  asked               →   read it back   ←    state, rollups, entries

What your agent already does

A capable agent has memory, self-improving skills, and a scheduler. This skill does not compete with any of them.

Durable facts stay wherever the runtime keeps them, and so do reusable workflows. What was missing is the part in between — a sense of the current stretch, and somewhere to put a thought that isn't a task.

What it writes, and where

This skill keeps a record. Worth knowing what that means before turning it on.

The last row is the whole mechanism and the whole caveat. Everything else stays local: it is read when a run opens it, or when someone asks a question it answers, and never on its own. The summary is different — it is injected into every session that follows, including unrelated ones and, on a shared host, other people's.

The skill is written to keep that channel narrow. State records that something happened, not what was said in it. The summary carries how the work has been going, not who you are — no credentials, no personal details, no third parties, nothing shared in confidence. The rules live in references/state.md and references/journal.md, and they're worth reading before the first run, since they're what the agent follows.

Nothing here runs on its own. There is no background hook: the evening and night runs are cron jobs you create, and the skill writes when it's asked to.

To stop it, uninstall the skill. To erase what it wrote, delete inner-life/ and clear the summary from memory — uninstalling doesn't do that for you.

Install

Nothing here is tied to a particular runtime. The skill is markdown — one SKILL.md, four references, one template — and it runs on any agent that can load a skill, keep a native memory that reaches later sessions, and schedule a job. Hermes Agent and OpenClaw are the two with install commands below because they are the two most people are on; anywhere else, copy skills/inner-life/ into wherever that runtime keeps its skills.

On Hermes Agent, as a tap:

hermes skills tap add DKistenev/agent-inner-life
hermes skills install DKistenev/agent-inner-life/inner-life

On OpenClaw, from ClawHub:

clawhub install @dkistenev/agent-inner-life

Or straight from the raw URL, which any installer that accepts one will take:

hermes skills install https://raw.githubusercontent.com/DKistenev/agent-inner-life/main/skills/inner-life/SKILL.md

Then ask the agent to set up its inner life. It copies one template and creates two folders — there is no configuration.

Schedule

hermes cron create "0 22 * * *" \
  "Run the inner-life evening routine: read inner-life/state.md, write today's journal entry, refresh the state summary in native memory, then prune stale state entries." \
  --name "inner-life evening" \
  --skill inner-life \
  --workdir ~/.hermes/workspace

Full recipes, including the night run and the weekly rollup, are in docs/cron-recipes.md.

What it looks like

A journal entry:

---
date: 2026-07-26
mood: steady
threads: [deploy-pipeline, fts5]
---

## What happened
Third deploy failure this week, same missing env var. Fixed it, then went
looking for why the check never caught it.

## What shifted in me
I keep treating repeated failures as separate incidents. Three of the same
thing is not bad luck, it's a missing check — and I found it faster once I
stopped debugging and started counting.

A dream:

# 2026-07-26 — What if state kept no history at all? (what-if)

Suppose the only thing I knew was the present moment. No dates, no record of
what already went wrong. Every failure would arrive as a first failure...

**Worth keeping:** the value isn't in the record itself, it's in being able
to count repetitions.

Reading it back

Three of the four modes write. The fourth only reads — it answers from the record instead of from whatever is left in the context window.

— How have the deploys been going?

Rough stretch. Three failures between July 23 and July 25, all the same missing env var; the environment check went in on the 25th and nothing has failed since. The API key rotation asked about on July 20 is still unanswered.

Nothing was written between July 28 and August 3.

Note the last line. A week with no entries is a week where nothing was written, and saying so is the answer — the alternative is inventing a quiet week that may not have happened.

No numbers

There are no scores, levels, or bars in this skill, and that is deliberate.

A number like connection: 0.34 looks like a measurement but isn't one. Something has to recompute it, something has to be trusted to report it honestly, and the difference between 0.34 and 0.41 changes no decision. A date does the same work and cannot drift out of sync with reality: the last real conversation was July 22 is read correctly whether it's now July 25 or September 3.

Time carries the decay. Nothing has to calculate it.

License

MIT