Continuum Genesis is an MIT reference implementation of task-scoped AI memory.
This repository is the runnable base product: local runtime, shard format, JavaScript SDK, installable text cockpit, tests, and a small evaluation harness. Thesis writing and public interpretation are separated into blog/.
No API key is required for the local reference runtime.
npm install
npm test
npm run seed
npm run shard -- --query "What should the follow-up email remember?"
npm run eval:memory -- --jsonStart the local cockpit:
npm startOpen:
http://127.0.0.1:8787/
| Path | What to inspect |
|---|---|
packages/memory-runtime/src/server.js |
Local HTTP runtime and API routes |
packages/memory-runtime/src/cli.js |
Seed, list, and shard CLI |
packages/memory-runtime/src/storage.js |
File-backed memory store |
packages/shard-format/src/index.js |
Retrieval scoring and context-shard builder |
packages/sdk-js/src/index.js |
JavaScript client |
apps/text-cockpit/ |
Installable browser/PWA interface |
tests/runtime/ |
Runtime, shard, PWA, and memory-lift tests |
scripts/evaluate-memory-lift.mjs |
Small local reproduction harness |
See docs/reproduce.md for exact commands and what each one proves.
Genesis stores durable project facts, retrieves only what the current task needs, and hands a compact context shard to a model, agent, CLI, or workflow.
memory items -> local runtime -> retrieval pass -> context shard -> model or workflow
The reference selector is intentionally simple: keyword and recency scoring. That keeps the mechanics auditable while leaving room for semantic retrieval, embeddings, graph memory, or agent-specific adapters.
| Endpoint | Purpose |
|---|---|
GET /health |
Check runtime status |
GET /v0/memory |
List stored memory items |
POST /v0/memory |
Add a memory item |
POST /v0/shards |
Build a task-scoped context shard |
Useful commands:
npm run seed
npm run shard -- --query "What should the follow-up email remember?"
npm run eval:memory
npm run security:triplenpm run seed refuses to replace existing memory. Use npm run seed -- --force
only when you intentionally want to reset the local demo store.
The base product stays generic. Local instance data lives outside the product code:
npm run instance:create -- instance-001
npm run instance:start -- instance-001The instance JSON is the oracle for mission, runtime, isolation, and upgrade rules. It keeps user receipts, model context shards, and audit logs as separate surfaces. See docs/instances.md.
Benchmark results live in docs/benchmarks.md, with each number labeled as officially scored, vendor self-reported, or independently reproduced.
The short version: Genesis holds the same credible accuracy band as other long-memory systems while optimizing for token efficiency, local ownership, and auditability.
For the argument behind that reading:
| Section | Purpose |
|---|---|
| docs/ | Technical reference, reproduction, benchmarks, release checks |
| blog/thesis/ | Formal measurement interpretation |
| blog/notes/ | Shorter public explanations |
| blog/reflections/ | First-person or experiential writing |
Before every public push, run:
npm run security:tripleThe release gate runs unit tests, leak checks, and a release-shape scan for local paths, provider key patterns, private runtime markers, customer-data markers, and service-worker caching mistakes.
See SECURITY.md, docs/public-private-boundary.md, and docs/release-checklist.md.
- Plainsight - MIT, device-local document risk reader.
- Throughline - MIT public archive of EVAN's first public venture.
MIT. Copyright ClearFrameworks LLC.