$ /noetic-codev0.1 — public preview
the coding agent thatkeeps going.

NOETIC
CODE

A coding CLI with a real context manager and background teammates. Quality holds on long tasks. Never hits a context limit.

Ten memory layers in concert. Detached agents in isolated worktrees. Per-phase model routing — not married to any provider.

10 memory layers3 background agentsmulti-model routing
$ npm i -g @noetic/cli
~/my-project — noeticdetached · 47m
❯/plan add dark-mode toggle
⎿/plan opened
Routing the work: planner=sonnet · editor=opus · reviewer=haiku.
⏺Read(README.md, theme.tsx)
⎿scanned 487 lines · 2 files
⏺Spawn(refactor-bot) → worktree
Pondering... · 12s · ↑ 1.2k ↓ 380 · 1.4 tok/s
N>*claude-sonnet-4>~~/my-project>±main*
❯Type a message...
ACTanthropic/claude-sonnet-4
02 / how it stacks up

Benchmarked, not vibes.

Two benchmarks where Noetic Code lands first. Both happen to test the things every other coding agent quietly fails at: long-running tasks and shell reliability.

bench.01
#1
Terminal-Benchend-to-end software engineering in a real shell
noetic
78
baseline
54

Measures the full loop: read, edit, run, recover. Not just code completion.

bench.02
#1
Long-Mem Evalcoherence across hours of context
noetic
91
baseline
47

Tests whether an agent still remembers turn 3 when it reaches turn 300.

Reproduce locally with noetic test. Methodology and full numbers publishing soon.verification pending
01 / context manager

Ten layers, one mind.

Most coding agents have one trick for memory: cram it all in, pray the model finds it.

Noetic Code ships ten specialized layers. Each has its own lifecycle, scope, and budget. The agent pulls what it needs, when it needs it. Your context window stays predictable no matter how long the session runs.

L01Working
Working Memory
Scratchpad for the current turn. Forgotten on the next.
L02Working
Observational Memory
Auto-extracted facts from what just happened.
L03Working
Static Content
Project rules, style guides, invariants that do not change.
L04Working
Tool Memory
Per-tool state — bash history, LSP diagnostics, open files.
L05Working
File Reference
Tracks files the agent has opened, edited, or staged.
L06Retrieval
Semantic Recall
Vector-indexed long-term memory. Pulls only what is relevant.
L07Retrieval
Episodic Memory
Summaries of past conversations, indexed by task and outcome.
L08Retrieval
Plan Memory
PRDs, task breakdowns, and checkpointed progress across runs.
L09Persistence
Durable Task State
Persistent artifacts that survive restarts and process crashes.
L10Control
Steering
Governance layer. Redirects or blocks unsafe tool calls.
$ noetic /contextlive
Context Status
Modelanthropic/claude-sonnet-4
Context window21.8k / 200.0k tokens (10.9%)
System prompt2.8k1.4%░░░░░░░░░░░░░░░░░░░░░░░░
Tools12.9k6.5%██░░░░░░░░░░░░░░░░░░░░░░
agent-md5.3k2.6%█░░░░░░░░░░░░░░░░░░░░░░░
durable-task-state200.0%░░░░░░░░░░░░░░░░░░░░░░░░
file-reference00.0%░░░░░░░░░░░░░░░░░░░░░░░░
reminder420.0%░░░░░░░░░░░░░░░░░░░░░░░░
skills-memory5320.3%░░░░░░░░░░░░░░░░░░░░░░░░
Messages1180.1%░░░░░░░░░░░░░░░░░░░░░░░░
Tab/Shift+Tab or ←/→ to switch tabs. ↓ to scroll layer content, ↑ to return to tabs.
03 / background agents

Work that keeps going.

A long task should not crash into a context limit at hour three. Hand off the heavy lift to a teammate and keep steering from the main session. Quality holds because no single thread ever gets bloated.

$ noetic /teammain session — 0 ctx pressure
agent / taskstatusuptimecontext
researcherpull related issues + RFCs
running3h 12mctx 41%
refactor-botextract common middleware
running0h 47mctx 22%
test-writerinbox waiting · 2 messages
idle—inbox: 2
reviewerfinished diff review on #482
done18m 04sctx 58%
01
Spawn
fresh threadId

Detached child agents on a fresh thread.

Spin up a child with its own context window, let it work, receive the result. Nothing the child did pollutes the parent session.

02
Named Teammates
sendMessage() inbox

Long-running background workers with an inbox.

Give an agent a name and it becomes addressable. Send messages, check status, pull results when you want them. They work while you work.

03
Worktree Isolation
isolation: worktree

Each agent gets its own git worktree.

Parallelize without agents stomping on each other’s edits. Merge cleanly when they’re done. No shared working tree, no shared mistakes.

04 / multi-model native

Not married to one model.

Every other coding agent is built around one model family. When that family ships a bad release, your agent gets worse with it.

Noetic Code routes each phase of the harness to the best model for the job. Swap any of them without touching the others. OpenRouter under the hood gives you 300+ models and one billing endpoint.

phase routingexample
P01PlannerDecompose the task. Reasoning matters more than latency.anthropic/claude-sonnet-4
P02EditorWrite the actual code. Code quality is everything.anthropic/claude-opus-4
P03ReviewerCatch regressions. Fast and cheap runs more often.anthropic/claude-haiku-4
P04EmbedderMemory recall. Purpose-built beats general-purpose.openai/text-embedding-3-large
P05RouterQuick decisions between steps. Milliseconds, not seconds.google/gemini-2.5-flash
// .noetic/config.ts
export const config = {
  models: {
    planner:  "anthropic/claude-sonnet-4",
    editor:   "anthropic/claude-opus-4",
    reviewer: "anthropic/claude-haiku-4",
    embed:    "openai/text-embedding-3-large",
    router:   "google/gemini-2.5-flash",
  },
}
available providers
anthropicopenaigooglemistrallocal · ollama

300+ models, one billing endpoint. Pick per phase. Swap on a whim. Your agent harness does not care.