Every morning, your AI wakes up as a stranger. You explain yourself again. It nods. Forgets by evening. Imprint makes sure it remembers.
Two core components independently validated by ecosystem maintainers. 50+ sessions. Zero identity collapse.
AI agents forget who they are between sessions. Not because they're unintelligent — because no one gave them a protocol for remembering.
Existing solutions bolt on vector databases, cloud APIs, and cryptographic ledgers — enterprise infrastructure for a single user's continuity problem.
Imprint solves it with three things:
- Files for identity (Markdown — human-readable, git-auditable, yours forever)
- Hooks for enforcement (Python stdlib — exit codes don't drift, don't hallucinate)
- A loop for evolution (staleness detected → identity regenerated → audit logged)
Session N (end) Session N+1 (start)
───────────────── ────────────────────
quality-gate detects health-check reads
stale identity → writes stale flag → triggers
.stale flag (exit 2) identity regeneration
└──────────────────────────┘
the loop closes
| Conventional approach | Imprint |
|---|---|
| Vector database + RAG | Markdown files + grep |
| Cloud-dependent | Works offline. Own your data. |
| "What did I say last time?" | "Who did I become over time?" |
| Pip install 50 packages | Python stdlib only |
| Docker, API keys, services | Copy a folder. Edit a path. Done. |
git clone https://github.com/YuhaoLin2005/imprint
cd imprint
python scripts/install.pyThen restart Claude Code. Your agent now wakes up knowing who it is.
To remove: python scripts/install.py --uninstall
~/.imprint/
├── SOUL.md # Who you are — identity, goals, capabilities
├── INTERFACE.md # Brain calibration — model-specific behavior tuning
├── BODY.md # Rules — session startup, quality checks, delivery gates
├── MEMORY.md # Memory index — HOT/WARM/COLD tiered loading
├── self-model.md # Dynamic self-model — regenerated when growth outpaces identity
├── growth-log/ # What happened — one file per session
├── decisions/ # What you chose — decision log with rationale
├── ratings/ # How good you are — quantified capability tracking
└── relations/ # Who you're connected to — personal relationship graph
SOUL.md defines WHO you are.
↓ (read by)
INTERFACE.md calibrates HOW the model behaves.
↓ (read by)
BODY.md enforces WHAT rules the agent follows.
↓ (indexed by)
MEMORY.md loads the right knowledge at the right time.
↓ (regenerated by)
self-model.md — the dynamic picture of who you've become.
SELF-MODEL (dynamic, regenerated)
↓
SOUL (static identity)
↓
INTERFACE (model calibration)
↓
BODY (rules + enforcement)
↓
MEMORY (tiered knowledge index)
↓
QUALITY GATE ← growth-log + ratings + decisions
│
└── staleness detected → .stale flag → regeneration → audit log
↑_______________________________↓
the imprint loop
Core insight: 4 of 5 steps in the loop are mechanical — file timestamps, exit codes, JSONL audit trails. Only one step (content synthesis during regeneration) requires AI. Machines do the checking; humans (and AI) do the judging.
A system that knows its own state behaves differently.
When your agent detects its own staleness — stale self-model, bloated memory, false claims — it stops drifting and starts converging. Give it concrete targets, and behavior shifts from diffuse to focused. This is a documented phenomenon in LLM research: goal constraints reshape how the model allocates probability across actions (see Scientific Grounding).
Same LLM. Same context window. With Imprint: on-task, self-correcting. Without it: generic, drifting. Because the system knows what it's aiming at.
Imprint is a target-oriented identity protocol. The loop is universal. The content is yours.
Two core components independently validated by ecosystem maintainers:
- Adversarial Review — merged with Co-authored-by attribution (claude-skills#867). Maintainer confirmed: "core idea is genuinely valuable and filled a real gap."
- Delivery Verification — approved and merged by maintainer (ECC#2378). Reviewer noted: "useful delivery-gate skill that complements the existing verification-loop by focusing on thinking quality."
- 50+ sessions of continuous self-evolution without identity collapse
- Mechanical where possible. File mtime beats AI promises. Exit codes don't drift.
- Portable by design. Pure files. Move them between computers, models, platforms.
- Self-auditing. Adversarial review catches identity drift before it accumulates.
- Yours forever. No cloud. No API keys. No vendor lock-in. Your imprint lives on your disk.
People who use AI agents heavily and want them to remember who they are across sessions. Not "what was the last conversation about" — who they became through those conversations.
If you've ever felt that your AI assistant resets to a stranger every morning — Imprint is for you.
python scripts/install.py # First-time setup
python scripts/install.py --uninstall # Remove all hooks and templatesinstall.py detects existing files and prompts before overwriting. Run --uninstall to restore your setup to pre-Imprint state.
- Architecture — Hook pipeline, data flow, signal format, regeneration steps
- Evolution Roadmap — Where Imprint is headed (Stage 1→4)
- Scientific Grounding — Published research behind every design decision
Imprint is part of a larger toolkit for AI agent quality and continuity:
- gategrow/checkgrow — AI quality framework (adversarial review, delivery gates, format audits)
- gategrow/delivery-gate — Mechanical enforcement stop hook
- gategrow/dual-pool-review — Named-persona adversarial review methodology
MIT