Continuum

Persistent memory for AI coding assistants that refuses to mark work done until a shell command proves it.

Every developer who works with an AI coding assistant has said it: “I told you about this last week.” The assistant starts every session blank, and you spend the first twenty minutes re-explaining the project. Continuum is the memory layer that ends that, and it earns every claim it makes by proving it with a shell command.

Who it helps

Continuum is for solo founders, small teams, consultants, and AI-assisted builders who work in Claude Code, Cursor, or Claude Desktop and lose hours each session re-establishing context. It serves three customers from one engine, and the engine is identical across all three; only the configuration changes.

Dogfood

The VoiceCosmos dev team runs it on its own work every day. If it does not hold up for us, it does not ship.

Independent developers

The open-source release, and a hosted option, for solo builders and small teams who want persistent project memory without standing up infrastructure.

Tenant-scoped enterprise

The same engine embedded per tenant, such as the ARIA Voice OS for hotels. Isolated data, shared discipline.

What it is

Continuum is a persistent intelligence layer and Model Context Protocol server: the hippocampus for an AI system. It continuously aggregates five sources of project truth into one verifiable memory database.

Docs Local documentation in your project, indexed for retrieval.

Memory Observations carried forward from past AI sessions.

Feedback Human-in-the-loop corrections and rewards.

Git Version-control history, one observation per commit.

Transcripts Full AI session logs, privacy-filtered on write.

How it helps

Verify-then-dissolve

The defining discipline. A task in the live todo pipeline is not marked done until its verifyCommand exits 0. No green shell exit, no completion. The only AI memory layer that refuses to take its own word for it.

A measured ~2.85x fewer tokens

Instead of dumping a codebase into context, the AI works through 3-layer Progressive Disclosure: a cheap search index, then chronological timeline, then full content only for the records it explicitly asks for. Benchmarked on this repo’s own history, reproducible with one command (not a marketing number).

Zero cold starts

At session start the AI reads one briefing resource and immediately knows the current state, open commitments, and what shipped yesterday. No re-explaining.

Local-first and private

SQLite + FTS5 today, RuVector vector search next, behind one StorageBackend seam. A secret-scrubbing filter sits on the single write path. Your history stays on your machine by default.

The ~2.85x figure is the measured aggregate from scripts/benchmark-token-savings.mjs (up to 5.3x when an answer needs a single record). We publish the number the benchmark produces, not the number we wish it produced. Run it yourself.

The end-to-end journey

Initialize. One command scaffolds the database, no external configuration:
Terminal window
```
npx @number7even/continuum-cli init
```
Warm start. The next session opens against continuum://session/briefing. The AI knows the product state, open todos, and yesterday’s progress before you type a word.
Token-efficient collaboration. While working, the AI queries the 5-source memory through continuum_search_docs → continuum_timeline → continuum_get_observations, paying full token cost only for what it needs.
Proof-driven execution. When the AI claims a task is finished, Continuum runs the task’s verifyCommand. It is marked complete only if the command passes.
State sealing. At session end the engine writes an append-only, hash-sealed checkpoint, so today’s architectural decisions can be queried as fact tomorrow.

Start here

Overview What Continuum is and how it works from your terminal.

Architecture The engineering source of truth: storage seam, MCP surface, locked decisions.

Build plan The gated, phase-by-phase roadmap to the V1 launch and beyond.

Install from npm @number7even/continuum-cli — the published CLI.