English | 日本語
An experimental harness that turns an AI agent's experience into evidence-grounded external memory, then compares raw context against raw context plus memory. This is not model-weight training or an official implementation of the systems that inspired it.
Python 3.11+, from a local repository checkout:
python3 -m venv .venv
.venv/bin/python -m pip install .
.venv/bin/agent-sleep-lab demo --format textThe demo uses built-in synthetic fixtures, scripted responses and disposable
temporary stores; --format text opts into a readable summary. The default
demo (or demo --format json) still prints the full JSON report. Installation
may download packages; the installed demo runtime is offline. It accepts no data,
store or provider options and makes no live-model calls.
First-run guide: output, cleanup and the separate live-library path → Demo contract →
Conceptual flow of the synthetic runtime; each stage requires explicit invocation, not an automatic or deployed agent loop.
flowchart LR
E["Capture / SQLite evidence"] --> S["Light / Deep proposals"]
S --> A["Atomic adoption"]
S --> V["Independent verification / evaluation"]
E -->|evidence| V
V -->|required receipts| A
A --> R["Scoped retrieval"]
E -->|raw context| C["Paired task evaluation"]
R -->|memory| C
SQLite is the runtime source of truth; JSON/Markdown are interchange artifacts. Providers propose memories, not verification receipts or execution permissions. Component map and separate public-document path →
The Connected Recall experiment has its own architecture, dependency boundary and rollback plan. SQLite remains canonical; Relate is proposed as a rebuildable relationship index for evidence-backed recall. It is separate from the quick start above. See design and alternatives and current implementation status.
| Looking for… | Start here |
|---|---|
| Current capabilities, verification and open gaps | Implementation status — the sole current-status authority |
| Commands and their execution conditions | Runtime guide |
| One-document admission, offline Light/Deep handoff, review, reuse and deletion | Public-document lifecycle — a separate library/store |
| Explicit library-only Light/Deep model calls and host role separation | Model connection contract |
| Integration, evaluation, research or development checks | Documentation index |
Scripted demo scores and manual draft handoffs are not live-agent integration or proof of improvement. Dated trials describe their tested versions, not current instructions. Public fixtures stay synthetic; private transcripts, credentials, memories and evaluator answers stay out of the repository. Paid inference needs an explicit bounded caller choice; byte limits are not token or monetary caps.
Package metadata is UNLICENSED; no open-source license is granted. See pyproject.toml.