Documentation · v3.0.1

Memory that lasts.

Clude is an open implementation of Stanford's Generative Agents — memory streams, reflection, and dream cycles, wired for production on Supabase + pgvector, with a local-first SQLite backend for offline use.

Install

One command, zero config. Detects your email from git, creates your agent, installs the MCP server for any editor you have:

npx @clude/sdk setup

Full install guide: clude.io/install.

First run

After setup, your agent gets four MCP tools: store_memory, recall_memories, get_memory_stats, and find_clinamen. Memories live at ~/.clude/brain.db by default.

Verify with:

npx @clude/sdk doctor

Memory types

Clude stores five kinds of memory. Each decays at a different rate, faithful to how human memory actually works:

Type	What it holds	Decay / day
episodic	Raw conversational moments, events	7%
semantic	Consolidated facts, stable beliefs	2%
procedural	Routines, tone, formatting preferences	3%
self_model	Agent identity, values, boundaries	1%
introspective	Reflections generated during dream cycles	2%

Recall

Clude's recall is hybrid: vector similarity (cosine, 384-dim embeddings) plus keyword matching plus tag scoring, with a composite score weighted by importance and recency. The math:

score = (0.45 × similarity) + (0.25 × importance) + (0.15 × recency) + (0.15 × tag_overlap) × decay_factor

In practice this means: a highly-relevant memory that was stored 3 months ago can still beat a weakly-relevant one from yesterday, as long as it was important.

Progressive disclosure

Context windows are expensive. Use recallSummaries() for ~50-token previews, then hydrate() only the full content of the memories you actually need:

const previews = await brain.recallSummaries({ query }); // rank or filter previews... const full = await brain.hydrate(previews.slice(0, 3).map(p => p.id));

Dream cycles

Every night (or on-demand), Clude runs a dream cycle:

Consolidation. Extracts semantic facts from recent episodic memories.
Compaction. Summarises faded episodic clusters into single gists.
Reflection. Generates introspective memories from patterns.
Contradiction resolution. Finds conflicting semantic memories, resolves via an LLM, stores the resolution with a resolves link.
Emergence. Links related memories across time into a knowledge graph.

Trigger manually:

npx @clude/sdk dream

Decay

Memories don't just accumulate. They fade. Each memory has a decay_factor in [0,1] that declines over time based on its type. Access boosts it (up to 0.1 per recall). Faded memories drop below the recall threshold and eventually get compacted.

This is the single biggest quality-of-life difference vs. a flat vector database. Old irrelevant context doesn't poison recall forever.

Cortex client

The SDK works the same way regardless of backend:

import { Cortex } from '@clude/sdk'; // Hosted (zero setup) const brain = new Cortex({ hosted: { apiKey } }); // Self-hosted const brain = new Cortex({ supabase: { url, serviceKey }, encryption: { solanaSecretKey: myKeypair.secretKey }, }); // Local-only const brain = new Cortex({ local: { dbPath: '~/.clude/brain.db' } }); await brain.init();

MCP tools

When Clude's MCP server is connected to your editor, agents get these tools automatically:

store_memory(content, type, source)
recall_memories(query, limit?, type?)
get_memory_stats()
find_clinamen(query) — finds surprising or tangential memories via anti-similarity.

Modes

Clude detects its mode from the environment:

Mode	Trigger	Storage
Hosted	`CORTEX_API_KEY`	clude.io · zero setup
Self-hosted	`SUPABASE_URL` + `SUPABASE_SERVICE_KEY`	Your Supabase
Local	no env vars set	`~/.clude/brain.db`

Self-hosted

If you want to run Clude on your own infrastructure:

Supabase project with pgvector extension enabled
Schema: curl -fsSL https://clude.io/supabase-schema.sql | psql $SUPABASE_DB_URL
Optional: Voyage AI for embeddings (VOYAGE_API_KEY)
Optional: Anthropic for dream cycle reflection (ANTHROPIC_API_KEY)

Point the SDK at it:

new Cortex({ supabase: { url: process.env.SUPABASE_URL, serviceKey: process.env.SUPABASE_SERVICE_KEY, }, });

Full deployment guide in the GitHub repo.