IPAC is a framework for encoding a person's knowledge, decision-making patterns, voice, and operational context into structured, version-controlled files that AI agents can load and execute.
The goal is practical: make the operator's working context portable, inspectable, and reproducible across sessions and models — instead of trapped in one person's head or one chat history.
The name and structure came out of a long working session in March 2026 while building tooling around long-running AI agents. It's a label for a pattern several people are converging on independently; prior art and parallel approaches are welcome and tracked in ideas/.
- IPAC-MANIFESTO.md — the full write-up
- spec/ — file format specifications
- examples/ — reference implementation templates
- ideas/ — open questions and research notes
- Blog post — longer-form context
Infrastructure as Code shifted server management from hand-configured one-offs to reproducible, version-controlled definitions. IPAC borrows that pattern for the human context an agent needs to operate on someone's behalf.
| IaC concept | IPAC equivalent | Implementation |
|---|---|---|
| Terraform state | Long-term memory | MEMORY.md + ref-*.md |
| Config files | Identity and values | SOUL.md, IDENTITY.md |
| Environment variables | Credentials/endpoints | TOOLS.md |
| User data | Human context | USER.md |
| CI/CD pipeline | Behavioral learning | instincts/*.yaml |
| Monitoring/alerts | Proactive checks | HEARTBEAT.md + cron |
| Drift detection | Memory maintenance | Vectorization + daily logs |
| Runbooks | Operational skills | Skill files |
| Backup/DR | Knowledge persistence | Git + embeddings |
- Write it down. If it isn't in a file, it doesn't survive a session restart.
- Version everything. Git is fine for tracking how context evolves over time.
- Fast feedback loops. Treat behavioral rules like CI: catch regressions early.
- Detect drift. Schedule maintenance so stored context doesn't diverge from reality.
- Portability. Any reasonably capable model should be able to load the files and produce comparable behavior.
- Refine, don't accumulate. Prune files; don't just append.
- Operator Kit — scaffolding for IPAC repos (think
terraform init, but for operator context) - OpenClaw — AI gateway runtime (third-party; we just use it)
- QMD — local vector search for memory retrieval
- Memory canaries — diagnostic primitives for testing agent session isolation (ideas/memory-canaries.md)
- Trust encoding — the harder problem sitting underneath context encoding (ideas/trust-encoding.md)
- Session = Sprint — applying sprint-style structure to agent sessions (blog post)
Active research, not a finished spec. The reference implementation has been running for about 8 months. Critiques, prior art, and PRs welcome.
CC BY-SA 4.0 — share and adapt with attribution.
Built at Imogen Labs. Questions: michael@mjashley.com