Skip to content

v0.1.0a1 — detections-in SDK

Pre-release
Pre-release

Choose a tag to compare

View all tags
@wikieden wikieden released this 04 Jun 14:46
v0.1.0a1
6cbf6a2

First public alpha. Spatial memory layer for AI agents — detections-in, persistent, queryable 3D scene graph. Core install is numpy-only.

Highlights

  • Fusion arbiter: incremental dedup — two sightings of one object converge to a single node (deterministic).
  • Query: spatial / temporal / keyword + semantic via a BYO Encoder protocol (OpenClipEncoder in [clip]).
  • sqlite-vec ANN ([vec]) with linear fallback.
  • answer() via a BYO Verbalizer (OpenAI / Anthropic / Ollama).
  • Memory hygiene: decay(), forget(), resplit() (split detection).
  • serialize(format="prompt"/"json") for LLM hand-off; bench.recall_at_k eval harness.
  • RGB-D seam: PerceptionAdapter protocol + add_frame() (ConceptGraphs adapter is WIP, needs CUDA).
  • Viewer: spatialmem viz store.smem -o scene.html.

Install

pip install spatialmem        # once published to PyPI
# or from source:
pip install "git+https://github.com/wikieden/spatialmem.git"

Pre-1.0 — API may change between minor releases. 52 tests, CI on 3.10–3.12 × macOS/Linux. See CHANGELOG.md.

Assets 2
Loading