agent-interface
The CLI and code contract your agent calls to init, ingest, recall and ask.
PROJECT
positronic-engram
Deterministic, auditable, tensor-grounded memory for LLM agents.
pip install git+https://github.com/ShingWong/positronic-engramCopyMost agent memory is a single-brain design: one embedding index decides what to store and what to retrieve. Engram treats curation as the hard problem instead. A salience gate scores surprise times goal-weight at encode time, so only what matters survives, and recall runs through four deterministic steps you can step through yourself.
MemoryEngine
A polytemporal MemoryEngine — predict → gate → encode | reinforce → consolidate → recall. Not a vector-store wrapper, not a summarization window.
Salience gating
Not everything gets stored: surprise × goal-weight decides what survives. Curation, not intelligence, is the hard problem of memory.
Polytemporal retention
Every event carries wall-clock, monotonic, subjective τ, and a fuzzy interval. Decay runs on Δτ — retention is a curve, not a TTL counter.
Tensor-grounded objects
Entities become evolving objects that persist across episodes. Activate · Reconstruct · Decay · Fuse — deterministic recall, no opaque magic.
Retention profiles
balanced / long_term / archival / short_term — same 55 messages, 78 weeks, different survival curves. Demotion, never deletion.
Speed
Store p95 0.7 ms (SQLite + FTS5 + RRF). One SQLite brain per session, offline and auditable.
Part of the ecosystem
Built alongside
Positronic is a small set of pieces that fit together — an engine, a seam, plugins and a network layer. Here is what sits around this project, and where each piece lives.
opencode plugin
The plugin that wires it into opencode with live auto-ingest on every turn.
serve
Expose a brain over HTTP, or federate several hosts into one logical memory.
research
Papers, architecture notes and the benchmark harness behind every figure quoted here.