neosian

The state layer for LLM agents: durable conversations, agent-curated memory, skills, a shared board and a ledger of who did what — on storage you own (a directory, your Postgres, or one small state process), for the agent you already use or the one you build.

Two doors, one store:

• Build an agent. An async-only, stateless Agent (tools, streaming, fallback, guardrails, structured output) that a Conversation wraps into a durable, memory-bearing thread.
• Give your agent state. Claude Code, Codex, Cursor or OpenCode get memory and skills over MCP and every session recorded through hooks — neosian mcp install --client claude-code, neosian record install.

A Claude Code session landing as a recorded turn that neosian audit names, and the next session opening on where we left off — no key of yours, no flags: the home and this project's scope.

Agents read this repository too: llms.txt is the machine front door, neosian docs prints the shipped docs pages from the wheel (version-true), and every shell verb has --json and exit tiers.

store = FileStore(home())                          # ~/.neosian, or your PostgresStore(dsn)
convo = Conversation(config, store=store, conversation_id="thread-829",
                     memory_scope=project_scope())  # user:<login>/proj:<dir>
response = await convo.send("Where did we leave off?")

Install

Requires Python >= 3.12. One package, everything in it — on PyPI as a pre-release until v1.0.0, so pin it:

uv add "neosian==1.0.0rc3"

The install brings the library with its three provider SDKs (OpenAI, Anthropic, Cerebras — xAI and Gemini ride the OpenAI wire), the neosian shell, the MCP server and client, the state process and OpenTelemetry spans: about 70 MB on disk, none of it loaded until used. Nothing in it needs an account or a server to start — Model.FAKE runs keylessly and FileStore is a directory. The one extra is the Postgres driver, uv add "neosian[postgres]==1.0.0rc3", for a PostgresStore against a server you already run; [all] is its alias, and the former [cli], [mcp], [otel] and [server] still resolve for one release and add nothing.

For a machine with nothing on it, the one-liner is scripts/install.sh: it says what it installs, finds uv or installs it from a pinned release, runs uv tool install neosian, checks the PATH and prints the registration command. It is the same two commands, spelled out beside it:

curl -fsS https://neosian.com/install | bash
# or, by hand:
curl -LsSf https://astral.sh/uv/install.sh | sh
uv tool install "neosian==1.0.0rc3"

To work on neosian itself, uv sync --all-groups then uv run neosian version (the Development section has the gates).

Quickstart

Keyless boot is an invariant: Model.FAKE is always available, so an install verifies with no account. The fake is scripted through the client_factory seam — the same seam the unit tier and the eval baseline use:

import asyncio

from neosian import Agent, AgentConfig, Message, Model, Role
from neosian.fake import FakeClient, FakeScript, FakeTurn

fake = FakeClient(FakeScript(turns=(FakeTurn(content="hello from the fake"),)))
config = AgentConfig(system_prompt="Be concise.", model=Model.FAKE,
                     client_factory=lambda _: fake)


async def main() -> None:
    response = await Agent(config=config).run(
        [Message(role=Role.USER, content="hello")], stream=False
    )
    print(response.message.content)


asyncio.run(main())

Set a key and name a real model (model=Model.CLAUDE_SONNET_5, the providers table below); tools are decorated functions whose signature is the schema (neosian docs tools). examples/ has one runnable file per feature, every one importing keylessly.

What is in the box

Each line is one page in the wheel — neosian docs <topic>.

• Conversation and memory (memory). History is an append-only log; log-projection compaction pages aged turns out of context and recall_turn re-hydrates any of them verbatim — paging, not deletion. Memory is file-school: small markdown documents with frontmatter under mounted scopes, read and written through one memory tool (view, create, str_replace, insert, delete, rename); reflection at the close distills the session into deliberate writes. Every write is a full-content version row with an actor; a memory_write frame on the stream carries {command, path, version} so a host renders "remembered X" with a working revert_memory undo.
• The shell and the ledger (cli). The six memory commands, the gardener (maintain), the operator verbs (versions, redact, revert), audit — what was done, by whom, when, on any substrate — and export/import, which move a store whole, history included.
• Skills (skills). A skill is a document under skills/<name> in a mount: versioned, curated by the mount flag, loaded by list_skills/load_skill, served over MCP as a prompt (a slash command in Claude Code).
• MCP, both directions (mcp). python -m neosian.mcp serves the store to any client over stdio; McpServer.stdio/http/in_process consumes any MCP server as agent tools — schema verbatim, is_error in-band, the approval gate and hooks unchanged.
• The agent you already use (agents). neosian record install spells the hooks that land a prompt-to-stop span as one turn by claude-code:<session> (Codex and OpenCode too), plus a sessions document; SessionStart prints the index and "where we left off" into the next window. A neosian Conversation on the same home shares the scope, so one audit names both.
• The state process (topology). neosian serve puts memory and conversations on a port (the shipped Dockerfile is the appliance) — reach, not capability; RemoteStore drops in where FileStore does, core install. Bearer tokens env-only, per-client when you want the ledger to know who wrote.
• The agent core (agent, tools). run(stream=True) yields ten frozen AgentEvent dataclasses a host relays over SSE; AgentHooks observe every turn, call and tool; ToolGateConfig routes every tool call through one approver, and no decision is a denial; capability- aware fallback, guardrails on any model, structured output, prompt caching, multimodal blocks on Anthropic.

Providers

API keys are read from environment variables; a provider is available when its key is set (asking for an unavailable model raises MissingAPIKeyError).

Every shipped row is earned by green dispatched runs of the memory baselines; membership is measured, never assumed (neosian docs baselines).

Storage

MemoryStore and ConversationStore are the contracts: async ABCs that own no connection, commit no transaction, issue no DDL. FileStore (a plain directory), PostgresStore (the postgres extra; schema by python -m neosian.schemas postgres | psql "$DSN") and RemoteStore (the state process's wire) implement both; a host may implement its own, kept honest by the shipped MemoryStoreContract / ConversationStoreContract conformance kits. One writer per FileStore root; many workers on one Postgres.

Evaluation

neosian eval suite.yaml runs YAML suites and exits nonzero on failure — an agent suite over variants × models × cases, and kind: memory, which scores store truth across scripted sessions on every transport. The shipped memory pack is all-green on models: [fake], and the same pack against the real providers produces the baselines page, fingerprint-gated so a prompt change without a recorded re-run fails make test.

Stability

v1.0.0 is deliberately not yet cut. When it is, it will carry the API stability promise: Agent, Conversation, MemoryStore, and the neosian.evaluation facade stable under SemVer, the ecosystem seams (scope grammar, token classes, integer micro-USD, event vocabulary, error codes) SemVer-guaranteed — a seam break only at a major — and the state process's wire (the twelve /v1/ store routes and their envelope, versioned by WIRE_VERSION) stable under the same promise: one promise covering library, seams, and wire. Until then the seams are append-only by convention, and error codes are already append-only forever. Consumers pin a release (neosian==X.Y.Z), never master.

Development

make install    # uv sync --locked --all-groups
make lint       # ruff + black --check + import-linter
make typecheck  # mypy --strict neosian tests examples
make test       # unit tier — the default gate, zero API keys
make size       # file-size gate (warn 300 / fail 500)
make test-external provider=anthropic file=~/path/to/creds  # real API calls
make test-postgres                                      # needs NEOSIAN_TEST_POSTGRES_DSN
make test-container                                     # both kits vs the container (docker)

The default gate needs no accounts; external tiers inject credentials value-blind. Sign off every commit (git commit -s): contributions are accepted under the Developer Certificate of Origin and licensed as the project is, Apache-2.0 — there is no CLA. Open an issue before a feature or a departure from documented behaviour; the public API is pinned by tests/unit/test_init.py, so an export change is a reviewed diff. SECURITY.md is where vulnerabilities go, never an issue.

Documents

• neosian docs <topic> — the shipped pages, from the wheel: quickstart, agent, tools, memory, skills, cli, mcp, agents, topology, baselines (the published per-provider memory numbers).
• SERVICES.md — every env key and what turning it off means.
• CHANGELOG.md — Keep a Changelog, one section per release.
• SECURITY.md — where to report, and the supported line.
• llms.txt — the machine-readable front door (byte-identical twin ships in the wheel).

Be kind and constructive in every project space; the maintainers read community@neosian.com.