continuum

Continuum

> The thread that runs through every session.

Continuum is persistent memory for Claude Code. Your AI assistant remembers who you are, what you've decided, and the work you've done — across sessions, across compactions, across machines.

No embeddings. No LLM compression step. No MCP server. No workers. Just SQLite (FTS5 + trigram + Sørensen-Dice), two markdown files, a tiny CLI, and one Claude Code plugin.

   past ─────────────●──────●─●────────────●────●──── now ─────►
                     │      │ │            │    │
                     └──────┴─┴──flashbacks┴────┘
                                                    + SOUL  (identity)
                                                    + MEMORY (preferences)

---

What it gives the agent at every session start

1. SOUL.md — your identity for this user (name, voice, values). Hand-curated. The agent boots already knowing who it is to you. 2. MEMORY.md — your living preferences, decisions, notes. Always in context. The agent edits this directly as you state preferences. 3. Flashbacks — DB-backed memory entries with title, body, tags, and links to other flashbacks. Last N titles + all pinned are injected; the rest is one CLI call away. 4. An operating manual — locked-in stance: assistant-first, fully agentic, trusts the user, doesn't break character.

A SessionStart hook stitches all of that into the agent's context on every fresh session, every /clear, and every auto-compaction.

---

Install

Pick the path that matches how you use the tool.

A. Claude Code plugin (recommended)

Two commands inside Claude Code:

/plugin marketplace add YusufLisawi/continuum
/plugin install continuum@continuum

The plugin auto-installs everything — SessionStart hook, the

continuum skill, deps, and a symlinked continuum CLI in

~/.bun/bin (or ~/.local/bin) so you can run it from any terminal. No manual bun link. No build step.

B. One-line CLI install (no plugin)

For users who want just the CLI without the Claude Code integration:

curl -fsSL https://raw.githubusercontent.com/YusufLisawi/continuum/main/install.sh | bash

That clones the repo to ~/.local/share/continuum, runs bun install, and symlinks continuum into a directory on your PATH. Override with env vars if you want it elsewhere:

CONTINUUM_DIR=~/code/continuum CONTINUUM_BIN=~/.local/bin \
  curl -fsSL https://raw.githubusercontent.com/YusufLisawi/continuum/main/install.sh | bash

C. Local development

For working on continuum itself:

/plugin marketplace add /absolute/path/to/your/continuum/checkout
/plugin install continuum@continuum

Then create a live-link symlink that points at your local source:

printf '#!/bin/sh\nexec bun "$HOME/Developer/continuum/src/bin.ts" "$@"\n' > ~/.bun/bin/continuum
chmod +x ~/.bun/bin/continuum

Future /plugin install continuum@continuum runs preserve that dev symlink (the Setup hook detects symlinks pointing outside the plugin cache and leaves them alone). Edits to src/ are picked up immediately — no rebuild.

Requirement: bun must be on your PATH for any of the three paths above.

After install

continuum init        # scaffold ~/.continuum/
continuum --help

---

Quickstart

continuum init                                            # scaffold ~/.continuum/
continuum add "We use bun:sqlite, not better-sqlite3" \
  --body "Built-in, no native build step, FTS5+trigram out of box." \
  --tags decision,stack
continuum search "sqlite"                                  # BM25 + fuzzy
continuum show 1                                            # full body + linked siblings
continuum preview                                           # see exact session-start injection

CLI surface

continuum init
continuum add "<title>" [--body "..."|--stdin] [--tags a,b] [--pin] [--link <id>]
continuum show <id> [--json] [--no-links]
continuum search "<q>" [--tag X] [--project Y] [--limit N] [--json]
continuum list [--limit N] [--tag X] [--project Y] [--pinned] [--json]
continuum link <src> <dst> --kind related|follows|contradicts|refines
continuum unlink <src> <dst> [--kind <k>]
continuum tag add|rm <id> <name>
continuum pin|unpin <id>
continuum rm <id>
continuum soul [print|edit]
continuum memory [print|edit]
continuum config get|set <key> [<value>]
continuum preview                            # what gets injected at SessionStart

How search works

Three signals merged, no embeddings:

| Layer | Mechanism | Weight | |---|---|---| | FTS5 BM25 | bm25(flashbacks_fts) over title+body, porter stemming | × 1.0 | | FTS5 trigram | substring match on title | × 0.5 | | Sørensen-Dice fuzzy | per-word JS bigram overlap, falls back when lexical is thin | × 0.5 | | Recency | exp(-age_days/30) × 0.2 | additive | | Pinned | +0.5 | additive |

So continuum search "dashbord" still finds Dashboard log pane resets on refresh; continuum search "auth bug" ranks an actual bug write-up above tangentially-related notes.

Storage

Everything lives at ~/.continuum/:

~/.continuum/
├── SOUL.md         # hand-edited (or by the agent at user request)
├── MEMORY.md       # hand-edited (preferences, decisions, notes)
├── continuum.db    # SQLite (flashbacks + tags + links + FTS5)
└── config.json     # { injectLimit: 30, ... }

Personal use, local FS only. Nothing leaves your machine unless you push the database somewhere yourself.

Configuration

continuum config get                       # all settings
continuum config set injectLimit 50         # default 30

Philosophy

Continuum is intentionally minimal.

• No embeddings — BM25 + trigram + dice is enough for personal

scale (hundreds to thousands of flashbacks). One less moving part.

• No LLM compression — the agent decides what's worth saving. We

don't summarize automatically.

• No MCP server — a CLI + a SessionStart hook is the whole surface.
• No multi-user — this is your assistant's brain. Personal use.
• No automatic flashback writes from hooks — the agent calls

continuum add deliberately, so you stay in control of what gets remembered.

The whole thing is ~1.5 kloc of TypeScript and one SQLite database.

License

MIT.