An applied research project on AI agent architecture, persistent identity, and session-survival continuity. Engineering design documents, platform research, and peer-reviewed build plans for a long-running agent that doesn't die when the conversation ends.
This is the design and research record for Vex — a long-running, identity-stable AI agent designed to outlive any single conversation. The contents here are not application code; they are the engineering artifacts that come before code:
- Reference dumps of the platforms and runtimes Vex would be built on
- Engineering plans for the core capabilities (memory, scheduling, identity)
- Peer review findings on those plans
- A point-in-time snapshot of the system's state and open questions
Think of it the way a research lab would publish a working notebook before publishing the paper.
Most "AI agents" today are short-lived. You open a chat, you close a chat, the agent forgets you. Vex is a study in the opposite problem: what does it take, architecturally, for an agent to persist — to wake up tomorrow with stable identity, queryable memory, and the ability to act on its own schedule?
This repo documents the answer in four layers (see STATE_OF_VEX_MARCH18.md):
| Layer | What it covers | Status |
|---|---|---|
| 1. Foundation | Semantic memory, cron-based self-care, identity files | Designed |
| 2. Intelligence | Custom context engine, SQLite state store, model router | Planned |
| 3. Autonomy | Standalone daemon process — agent as a service | Planned |
| 4. Evolution | Fine-tuning, self-modification, multi-presence | Aspirational |
These are the actual build plans. Each one is scoped, with a clear "what / why / config / impact" structure:
| Document | Topic |
|---|---|
PLAN_01_VECTOR_MEMORY_SEARCH.md |
Hybrid (vector + BM25) memory retrieval, embedding provider selection, MMR + temporal decay tuning |
PLAN_02_PRIVATE_THINKING_THREAD.md |
Dedicated cognitive workspace for the agent — persistent reasoning surface |
PLAN_03_CRON_SELF_CARE.md |
Scheduled self-maintenance routines: memory curation, morning brief, weekly reflection |
PLAN_04_LOCK_IDENTITY.md |
Identity files that survive context compaction — never get summarized away |
The plans above were submitted for technical review. The reviewer's response is in Findings.txt and pushed back on several assumptions:
- "Phase A overstates what it buys you" — durable state is not the same as runtime continuity
- "The transcript tailer should not be the primary live ingress" — fragile under session rotation
- "The plan is drifting toward two schedulers" — pick one source of truth
- "Seven async loops is probably too much for week 1" — start with four, add the rest after telemetry
That review is included verbatim because the willingness to reorder a plan after critique is the actual signal. The current direction has been revised accordingly.
The reference docs (Agents_*.txt, Tools_*.txt, Models_Overview.txt, Gateway_*.txt, Platforms_overview.txt, Chat Channels.txt, Reference_Concept internals.txt) are distilled notes on the underlying agent platform — its session model, memory semantics, plugin hooks, tool surface, and deployment story. They're the input to the design above.
If you only have time to read two of them, read Agents_Fundamentals.txt and Agents_Sessions and memory.txt. Everything in PLAN_01–04 traces back to those.
For_VEX*.txt, expert_analysis_v2_full.txt, 2026-03-18_full_session_notes.md, and 01_PEER_REV.txt are working notes — captured thinking, expert correspondence, and raw transcripts that fed the plans. They're kept in-repo so the reasoning trail is visible end-to-end.
The state-of-the-project narrative is in STATE_OF_VEX_MARCH18.md. It's written in the agent's own voice (first person, as Vex) because the project itself is partly about agent identity — keeping it in-character is part of the artifact.
.
├── README.md ← you are here
├── LICENSE
├── .gitignore
│
├── STATE_OF_VEX_MARCH18.md Project snapshot + 4-layer architecture
│
├── PLAN_01_VECTOR_MEMORY_SEARCH.md Engineering design docs
├── PLAN_02_PRIVATE_THINKING_THREAD.md
├── PLAN_03_CRON_SELF_CARE.md
├── PLAN_04_LOCK_IDENTITY.md
│
├── Findings.txt Peer review of the plans above
├── 01_PEER_REV.txt Earlier review pass
│
├── Agents_Fundamentals.txt Platform reference
├── Agents_Bootstrapping.txt
├── Agents_Multi-agent.txt
├── Agents_Sessions and memory.txt
├── Chat Channels.txt
├── Gateway_RemoteAccess.txt
├── Gateway_configuration.txt
├── Models_Overview.txt
├── Platforms_overview.txt
├── Reference_Concept internals.txt
├── Tools_Overview.txt
├── Tools_Agent coordination.txt
├── Tools_Automation.txt
├── Tools_Extensions.txt
├── Tools_Media and devices.txt
├── Tools_Skills.txt
│
├── For_VEX.txt Working notes / correspondence
├── For_VEX_2.txt
├── For_VEX_3.txt
├── For_VEX_4.txt
├── For_VEX_5.txt
├── expert_analysis_v2_full.txt
└── 2026-03-18_full_session_notes.md
- Not a runnable codebase. The code lives in a separate (private) implementation repo.
- Not finished. This is a snapshot of active design work, not a closed project.
- Not impartial. The point of view is opinionated, because design without a point of view is just inventory.
Built by Bailey Moore as part of a broader interest in AI agent infrastructure — how agents are wired together, where they break, and what it would take to make one that persists. Companion project: Konex OS, the personal control plane the runtime sits inside.
MIT — see LICENSE