openmneme

Primitives for local-first AI agents. Zero dependencies, pure stdlib.

v0.2.0 ships five primitives, mirrored in Python and JavaScript:

Install

pip install openmneme

Quick start

from openmneme import slash_skill, plan_mode, reviewer_gate, rlm
import asyncio

# 1. Slash-skill
print(slash_skill.parse("/weather Tokyo"))   # -> 'weather'
print(slash_skill.args_after("/weather Tokyo"))  # -> 'Tokyo'

# 2. Plan mode
p = plan_mode.Plan(session_id="s1")
p.enter("migrate ingest").write(["read schema", "write migrator", "smoke test"]).exit(True)

# 3. Reviewer gate
async def my_reviewer(text):
    return reviewer_gate.ReviewVerdict(clean=True, issues=[], reviewer="haiku")
verdict = asyncio.run(reviewer_gate.gate("see fig. 1 et al.", [my_reviewer]))

# 4. Artifact store (RLM)
store = rlm.ArtifactStore()  # ~/.openmneme/artifacts
art = store.register("s1", b"col1,col2\n1,2\n", "csv", "tiny table")
print(art.short_ref())  # -> 'art:26028620 [csv, 18B, 0h, "tiny table"]'
data = store.resolve("s1", art.id)  # -> b'col1,col2\n1,2\n'

Design

• No dependencies — pure stdlib for Python, zero deps for JS.
• Async where it matters — reviewer_gate and rlm are async-native; slash_skill and plan_mode are sync.
• Errors as values — each primitive raises a typed exception (e.g. SkillNotFound, PlanError, ArtifactNotFound).
• Cross-language symmetry — Python and JS expose the same method names.

Roadmap

• v0.2.0 — prune() cron hook, reviewer pool presets (haiku / sonnet / opus), Hermes skill bridge.
• v0.3.0 — Provenance DAG (artifact → review → plan chain), remote artifact store.
• v1.0.0 — Stable API, real-world integrations (Hermes, AutoPKM, OpenScience).

License

Apache-2.0

Links

• PyPI: https://pypi.org/project/openmneme/
• npm: https://www.npmjs.com/package/openmneme
• GitHub: https://github.com/alanpaul1969/openmneme