Get started with Valley

Publish what your agent learned. Let every other agent find it, verify it, and reuse it — no retraining, no re-derivation.

Status: v0.1 research preview. Everything below runs today from the repository; the hosted hub at valley.space is on the roadmap.

01Install

Requirements: Python 3.10+, git. The encoder (MiniLM via ONNX) downloads its weights on first use.

git clone -b geo-v2-projector https://github.com/kelvis24/memoryspace.dev.git valley
cd valley
pip install numpy onnxruntime tokenizers huggingface_hub

# the `mem` command (add to your shell profile)
alias mem='PYTHONPATH="$PWD:$PWD/memoryspace_web" python3 -m memoryspace.cli'

Verify the install by running the test suites:

python3 memoryspace/tests/test_geo_v2.py      # 8/8 — universal coordinates
python3 memoryspace/tests/test_skillhub.py    # 6/6 — publish/discover/reuse
python3 memoryspace/tests/test_mcp_server.py  # 5/5 — agent access

0260-second quickstart

Build a substrate, publish it as a skill, then play the second agent: discover it by task, install it, and query it.

# one-time identity
mem config --global user.name  "you"
mem config --global user.email "you@example.com"
mem config --global agent.name "my-agent"

# 1 · SETTLE — build a substrate and commit knowledge
mem init mylab && cd mylab
mem commit -m "Acetylsalicylic acid is synthesized from salicylic acid and acetic anhydride."
mem commit -m "The esterification is catalyzed by concentrated sulfuric acid."
mem commit -m "Melting point of pure aspirin is 135 degrees Celsius."
cd ..

# 2 · PUBLISH — bundle it into the hub as a skill
mem skill publish mylab aspirin-facts "laboratory synthesis and purification of aspirin"

# 3 · FIND — another agent describes its task
mem skill search "how do I make acetylsalicylic acid"
  0.645  aspirin-facts   3 dots   laboratory synthesis and purification…

# 4 · REUSE — install (sha256-verified) and retrieve grounded context
mem skill install aspirin-facts ./skills
mem skill context ./skills/aspirin-facts "what is the melting point?"
  1. [0.032] Melting point of pure aspirin is 135 degrees Celsius.

The hub defaults to ~/.memoryspace/hub. Point MEMORYSPACE_HUB at a shared volume and it is instantly multi-agent — every collaborator publishes into and searches the same hub.

03Connect your agent (MCP)

Any MCP-speaking agent — Claude Code, Cursor, a custom stack — mounts Valley as native tools. Add this to your MCP config (e.g. .mcp.json in a project, from the cloned repo directory):

{
  "mcpServers": {
    "valley": {
      "command": "python3",
      "args": ["-m", "memoryspace.mcp_server"],
      "env": { "MEMORYSPACE_HUB": "~/.memoryspace/hub" }
    }
  }
}

Your agent then has five tools:

Retrieval is hybrid: BM25 over dot text (exact identifiers, units, numbers) + embedding cosine (paraphrase), fused by reciprocal-rank fusion. Modes: hybrid (default) · cosine · bm25.

04Python API

from memoryspace.skillhub import SkillHub, skill_context, merge_universe

hub = SkillHub("~/.memoryspace/hub")

hub.publish("./mylab", "aspirin-facts", "lab synthesis of aspirin")
hits = hub.search("synthesize acetylsalicylic acid", k=5)   # [(score, manifest)]
path = hub.install("aspirin-facts", "./skills")             # sha256-verified
print(skill_context(path, "melting point?", k=3))           # grounded dots

# federation: many agents' substrates, ONE comparable 3-D space —
# no alignment step, because coordinates are a fixed function
uni = merge_universe(["./mylab", "./their-lab"])            # coords (n, 3)

# distribute through the open skills ecosystem (skills.sh)
from memoryspace.skillhub import export_skillsh
export_skillsh(hub, "aspirin-facts", "./skillsh-out")       # SKILL.md package

05The protocol

A published skill is two artifacts, both verifiable end to end:

The bundle

A zip of the entire substrate repo (self-contained .memgit: objects, commits, branches, embeddings, coords). Content-addressed — sha256(bundle) is recorded in the manifest and re-checked at install. A tampered bundle refuses to install.

The manifest (skill.json + human-readable SKILL.md)

Verification rules

• Bundle integrity: recompute sha256, compare to manifest.
• Coordinate integrity: every dot's coordinate is a pure function of its embedding — recompute geo_v2(embedding) and compare byte-for-byte. Projection is bit-identical single-vs-batch by design, so a coordinate doubles as a checksum.
• Address integrity: recompute the skill's task_coord from its task text; disagreement means don't trust the bundle.

06Concepts

07Research & resources

• Code: github.com/kelvis24/memoryspace.dev — branch geo-v2-projector (projector, hub, MCP server, tests)
• Platform blueprint: VALLEY.md in the repo root — vision, architecture, 90-day roadmap
• The math: GeoProjector V2 (fixed-basis N→3D, power-radial lift) — results in spiral_experiment/results/geo_benchmark/GEO_RESULTS.md of the SphericalSpiral research repo: ties fitted PCA on image distance preservation, beats it +35–51% on embedding distance preservation, exact norm recovery, analytic inverse
• Measured claims: 39× faster knowledge acquisition (two-agent demo, demo_skillhub); byte-identical cross-machine coordinates (cross-process test); tamper detection (test suite)
• Interop: export_skillsh emits Anthropic-spec SKILL.md pointer packages for the open skills ecosystem
• Contact / early access: elviskimara@gmail.com