agent-sleep
/agent-sleep
Porządkuje pamięć agenta kodującego po sesji: konsolidacja, higiena pamięci i kontrola dryfu, zawsze poza trybem interaktywnym.
- Repozytorium
- agent-sleep ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 03.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 4
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 03.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Porządkuje pamięć agenta kodującego po sesji: konsolidacja, higiena pamięci i kontrola dryfu, zawsze poza trybem interaktywnym.
Oryginalny opis (EN)
Sleep-cycle memory management for coding agents. Use when consolidating a finished session into memory, running daily memory hygiene or drift checks, prestudying a repo before the next session, reconciling accumulated memory, rebuilding memory from source, or installing the wake/sleep triggers on a new machine or platform. Modes are offline by design — never run consolidation inline during an interactive session.
Jak działa
- The invariants
- Memory file format
- Running the modes
- What this is not
Podgląd SKILL.md
---
name: agent-sleep
description: >-
Sleep-cycle memory management for coding agents. Use when consolidating a
finished session into memory, running daily memory hygiene or drift checks,
prestudying a repo before the next session, reconciling accumulated memory,
rebuilding memory from source, or installing the wake/sleep triggers on a
new machine or platform. Modes are offline by design — never run
consolidation inline during an interactive session.
---
# agent-sleep
A memory architecture for coding agents, built around one split: sessions
work, offline phases remember. The agent that does the task should not be the
agent doing memory bookkeeping in the same breath.
Six modes, each with a trigger:
| Mode | Trigger | What runs |
| --- | --- | --- |
| consolidate | session end (off-peak) | extract → verify vs raw source → promote |
| check | daily, scripts only | TTL expiry, pointer re-validation, founding tests |
| reconcile | N new memories, or a contradiction flag | dedupe, prune, scoped rewrites |
| prestudy | idle (off-peak) | re-derive repo state into warm-start memory |
| rebuild | re-grounding costs more than maintaining | regenerate memory from source |
| rollback | drift alarm, or any bad write | restore last-known-good from git |
## The invariants
These exist because specific, measured failure modes were documented
(see README for the evidence). Breaking one is a bug even if nothing
visibly breaks today.
1. **Never summarize a summary.** Extraction reads the raw transcript and the
repo. Existing memory is never an input to consolidation.
2. **Never write the charter.** `AGENTS.md` (or platform equivalent) and skill
cores are re-loaded fresh each session and live outside every write path
this skill offers. If a mode seems to need charter edits, stop and report.
3. **Admit less.** A candidate memory enters only if it passed verification
against a source and carries a scope. When in doubt, drop it.
4. **Every write is a git commit.** One writer at a time, one commit per
promotion, via `scripts/snapshot.sh`. No exceptions, no batch rewrites
outside reconcile.
5. **Check daily, rewrite on trigger.** Mechanical checks are cheap and
frequent; semantic rewrites are rare and batched. Each rewrite is a
compression event — spend them.
## Memory file format
Every memory is one file with frontmatter (see `templates/memory-file.md`):
```yaml
name: short-kebab-slug
description: one line, used for recall decisions
metadata:
type: user | feedback | project | reference
scope: when this applies — and when it does not
source: /path/or/URL/or session:<id>
verify: optional command that must exit 0
ttl_days: 30
last_validated: 2026-10-28
```
An index file (`MEMORY.md`) lists one line per memory and nothing else.
The index routes recall; the files carry the facts.
## Running the modes
Scripts are plain bash; the only dependencies are git, curl, and a POSIX
shell. Model-driven steps use the prompt templates in `templates/prompts/`
with the strongest model available — consolidation quality is the whole game.
scripts/snapshot.sh <memory-dir> [message] # commit memory state
scripts/rollback.sh <memory-dir> [N] # restore N commits back
scripts/verify-memory.sh <memory-dir> # re-validate every entry
scripts/prune-ttl.sh <memory-dir> [--apply] # list/move expired entries
scripts/run-fixtures.sh <fixtures-dir> # founding tests, exit 1 on drift
scripts/install-triggers.sh # platform wiring
Platform-specific trigger wiring (hooks, idle queues, cron) lives in
`adapters/` — one file per platform, starting with ZCode, Claude Code,
OpenCode, and a generic cron adapter that covers the rest.
## What this is not
It is not parametric learning. Everything lives in token space — files a
harness reads. That ceiling is real (weight-space consolidation outperforms
it in the literature), and the payoff is that memory survives every model
swap underneath.
Pliki w paczce (4)
- .gitignore 1 KB
- LICENSE 1 KB
- README.md 5 KB
- SKILL.md 4 KB
Wersje
W bazie nie ma innych skilli o nazwie agent-sleep.
Migawka bazy z 13.11.2026 · wersja paczki 2e6a845d2d25 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 95sleep
Przeprowadza cykl pamięci: zapisuje obserwacje, łączy je w kandydatów i awansuje tylko to, co przeszło pomiar powtórką. - 100sleep
Prowadzi pamięć agenta przez cykl T0, T1 i T2: zapisuje obserwacje, sprawdza kandydatów i awansuje zatwierdzone wpisy. - 100skillopt-sleep
Uruchamia nocny cykl samodoskonalenia agenta na podstawie minionych sesji, konsolidując pamięć i skille. - 100skillopt-sleep
Uruchamia silnik skillopt_sleep, dzięki któremu agent Claude uczy się na dawnych sesjach: zbiera je, wydobywa powtarzalne zadania, konsoliduje pamięć i skille oraz planuje nocny cykl samodoskonalenia offline. - 96sleep
Uruchamia przebieg snu LR Memory: zapisuje reguły, poprawki i fakty z rozmów w pamięci, na żądanie użytkownika. - 100skillopt-sleep
Pozwala agentowi Claude uczyć się z dotychczasowego użycia w nocnym cyklu „sleep”: zbiera sesje, wyciąga powtarzające się wzorce i konsoliduje pamięć oraz skille.