Squish - Memory Infrastructure for AI Agents

SDK, MCP server, and CLI for persistent agent memory. Local-first. Zero dependencies.

Squish is memory infrastructure for AI agents. Use the TypeScript SDK to build memory-powered apps, the MCP server to plug into any compatible agent, or the CLI for quick operations. Local-first with optional Cloud sync.

If Squish saves your agent's memory, give it a star -- it helps other builders find it.

Get Started in 30 Seconds

npm install -g squish-memory && squish install --all

That is it. Squish installs the CLI, starts the MCP server, and configures hooks for every coding agent it finds on your machine. No API keys. No config files. No Docker.

Build on Squish

The @squish/sdk is the recommended way to build memory-powered AI applications:

import { SquishClient } from '@squish/sdk';
const client = new SquishClient({ dataDir: './memory' });
await client.remember('key decision', { type: 'decision' });
const context = await client.getContext();

SDK Documentation | API Reference

Pick Your Agent

Squish works with any MCP-compatible agent. Choose yours for a tailored quick start:

Claude Code

npm install -g squish-memory && squish install --all

Squish detects Claude Code and adds plugin hooks automatically. Your next session starts with full memory context. To verify:

squish context    # See what your agent remembers
squish status --stats  # Check memory health

Codex CLI (OpenAI)

Add Squish to your Codex MCP config:

{
  "mcpServers": {
    "squish": {
      "command": "squish-mcp",
      "args": ["--http", "--port", "8767"]
    }
  }
}

Codex now has persistent memory across sessions. Ask it "what did we decide about the database?" and it will recall your past decisions.

Cursor / Windsurf / Cline

Add the same MCP server block to your editor's MCP settings. One memory server, shared across all your editors and CLI agents.

OpenCode

squish install --all

OpenCode gets both MCP tools and auto-capture hooks. Decisions, constraints, and preferences are captured as you work.

Any MCP Client

{
  "mcpServers": {
    "squish": {
      "command": "squish-mcp",
      "args": ["--http", "--port", "8767"],
      "env": {
        "SQUISH_DB_PATH": "./squish-data"
      }
    }
  }
}

What Just Happened

After install, Squish runs in the background. Here is what it does:

Captures -- As you work, Squish watches for decisions, constraints, preferences, and context. It filters noise and stores what matters.
Ingests -- Drop files into the inbox directory. Images, audio, video, and documents are automatically extracted, described, and stored as searchable memories.
Stores -- Memories go into a local SQLite database with AES-256-GCM encryption. Nothing leaves your machine.
Retrieves -- When your agent starts a new session, Squish injects only the relevant memories (50-200 tokens, not 2,000).
Decays -- Old, low-value memories fade automatically. Your agent stays focused on what matters now.

squish remember "We chose PostgreSQL for the main datastore" --type decision
squish recall "database decisions"
squish sessions search "postgres migration"

Works with Every Agent

Agent	Integration	Auto-Capture
Claude Code	MCP server + plugin	Yes
Codex CLI	MCP server	No
Cursor	MCP server	No
GitHub Copilot	MCP server	No
Gemini CLI	MCP server	No
OpenCode	MCP server + hooks	Yes
Cline	MCP server	No
Goose	MCP server	No
Windsurf	MCP server	No
Roo Code	MCP server	No
Claude Desktop	MCP server	No
Aider	MCP server	No

One memory server. Shared across all of them.

The SDK (@squish/sdk) works with any TypeScript or JavaScript environment -- not just coding agents. Build memory into web apps, CLI tools, backend services, or custom agent frameworks.

Why Squish

Most memory tools need a second LLM for embeddings and retrieval. That means extra API costs, latency, and infrastructure you have to manage.

Squish uses local embeddings by default. Zero LLM dependency. 1-5ms latency. $0 runtime cost in local mode.

Comparison

Feature	Squish	CLAUDE.md	agentmemory	mem0	Letta	Zep
SDK available	Yes (TypeScript)	No	No	No	No	No
Auto-capture	Yes (hooks)	Manual	Yes (12 hooks)	Manual API	No	No
Local embeddings	Yes (default)	N/A	Yes	No (cloud)	No	No
External DB required	No (SQLite)	No	Yes (iii-engine)	Yes (Qdrant)	Yes (Postgres)	Yes (Postgres)
MCP tools	15	0	53	9	0	0
Knowledge graph	Yes	No	Yes	No	Yes	Yes
Cross-agent sync	Yes (Cloud)	No	No	API-based	API-based	API-based
Price	Free local / $9/mo cloud	Free	Free	$249/mo Pro	Free (self-hosted)	Free (self-hosted)
Setup time	30 seconds	5 minutes	15 minutes	30 minutes	30+ minutes	30+ minutes
Data ownership	Full (local SQLite)	Git repo	External DB	Cloud vendor	External DB	External DB

What Makes Squish Different

Squish is the only tool that combines all of these in a single package:

SDK-first architecture -- @squish/sdk for building memory-powered apps in any TypeScript/JavaScript environment
Local-first by default -- SQLite storage, TF-IDF embeddings, zero API keys needed
MCP-native -- Works with every MCP-compatible agent out of the box
Knowledge graph -- Reinforced relationships between memories, not just flat storage
Decay system -- Old, low-value memories fade automatically so agents stay focused
Hybrid search -- BM25 keyword + semantic similarity with RRF fusion
Multimodal ingestion -- 27+ file types: images, audio, video, documents
Zero external dependencies -- No Qdrant, no Postgres, no vector DB required in local mode

Core Concepts

Concept	What It Is
Recall	Durable memory -- decisions, preferences, constraints
Sessions	Evidence from past agent runs
Pinned	Stable facts that do not decay
Beliefs	Passive model of user/project
Strategies	Active operating rules
Media Memories	Ingested images, audio, video, and documents with extracted text
LLM Consolidation	Cross-connection finding via LLM-powered knowledge analysis
Decay	Stale weak traces fade automatically
Graph	Reinforced relationships from usage

Features

Memory Intelligence

Auto-captures decisions, constraints, and preferences as you work
Restores relevant context when an agent restarts
Handles contradictions and temporal facts with expiration
Graph-boosted retrieval connects related memories across sessions
Contradiction detection flags conflicting information
Temporal reasoning tracks when facts were true vs. now
Confidence scoring adjusts memory relevance over time
Decay system automatically ages low-value memories

Multimodal Memory

Ingest images, audio, video, and documents into searchable memories
Automatic text extraction via OCR, speech-to-text, and document parsing
27+ supported file types: JPEG, PNG, GIF, WebP, TIFF, HEIC, MP3, WAV, OGG, FLAC, M4A, MP4, WebM, AVI, MOV, MKV, PDF, DOCX, XLSX, PPTX, TXT, MD, CSV, JSON, XML, YAML, HTML, RTF
File watcher for automatic inbox monitoring and ingestion
LLM-generated descriptions for each ingested file
Cross-connection finding via LLM consolidation across memory clusters

Session Search

Search previous Claude Code, Codex, and OpenCode sessions
Find related sessions by project path or file overlap
Inspect past decisions, errors, and commands as evidence
Separate from long-term memory -- raw session history, not distilled facts

Interfaces

CLI: squish remember, recall, forget, link, context, stats (via status), clean, pin, sessions, doctor, cloud
MCP Server: 15 tools for any MCP client -- remember, recall, forget, link, context, stats, inspect, skill, loadout, extract, feedback, places, sessions, tier, dedup (a 16th, squish_maintenance, appears when SQUISH_ENABLE_MAINTENANCE_TOOLS=true)
Web UI: Local dashboard at localhost:37777 for visualizing memories
Cloud Dashboard: Paid analytics and management at squishplugin.dev

Storage

SQLite (local, default) or Squish Cloud team workspaces
Hybrid retrieval: keyword + semantic similarity with RRF fusion
AES-256-GCM encryption for sensitive memories
Places routing: organize memories by project, feature, or context
Full-text search with BM25 ranking
Vector search with TF-IDF embeddings (768-dimensional)

Self-Hosting

Squish is local-first by design: the default mode IS self-hosted. Everything runs on your machine, no account required.

Storage: single encrypted SQLite file (AES-256-GCM for sensitive memories) at SQUISH_DATA_DIR (default ~/.squish)
Embeddings: local TF-IDF by default -- zero API keys, works fully offline. Optional providers (openai, ollama, google) if you want them
LLM: optional and off by default (SQUISH_LLM_ENABLED=false) -- every core feature works without one
Telemetry: none. Nothing leaves your machine in local mode
License: MIT. Fork it, embed it, ship it

Point SQUISH_DATA_DIR at any volume or backup target and you are done. See Environment Flags for tuning knobs.

Environment Flags

Every flag ships with a safe default; nothing here is required for local use. See .env.example for the full annotated list.

Flag	Default	What it does
`SQUISH_DATA_DIR`	`~/.squish`	Where the SQLite DB and assets live
`SQUISH_EMBEDDINGS_PROVIDER`	`local`	`local` / `openai` / `ollama` / `google` / `none` / `auto`
`SQUISH_LOCAL_BUNDLED_MODEL`	bundled MiniLM	Set `off` to pin deterministic TF-IDF (CI, offline evals)
`SQUISH_VECTOR_SCAN`	`recency`	Candidate selection: `full` scans every row (complete recall), `recency` sees the newest window
`SQUISH_SEARCH_BELIEFS`	`true`	Include belief records in hybrid retrieval
`SQUISH_ABSTAIN_BELOW`	unset	Recall-confidence floor; below it recall returns a `no_reliable_memory` verdict instead of weak hits
`SQUISH_SCORING_V2`	`true`	Serve v2 composite ranking (`SQUISH_SCORING_SHADOW=true` logs v2 alongside without serving)
Precision stack	see docs/v2-scoring.md	`SQUISH_RERANKER_ENABLED`, `SQUISH_QUERY_EXPANSION`, `SQUISH_TEMPORAL_VALIDITY`, `SQUISH_GRAPH_BOOST_LEGACY`, `SQUISH_MMR_ENABLED`, `SQUISH_CONTEXTUAL_RETRIEVAL`, `SQUISH_ENTITY_RETRIEVAL`
`SQUISH_GEOMETRY_CONSOLIDATION`	`true`	Geometry-aware consolidation guards (theta-prime split/consolidate decisions)
`SQUISH_DEDUP_AUTO`	off	When `true`, nightly dedup may execute high-threshold merges automatically (capped per run); otherwise proposals wait for review via `squish_dedup`
`SQUISH_ACL_ENFORCE`	`false`	Enforce visibility rules on memory reads
`SQUISH_LLM_ENABLED`	`false`	Enables optional LLM features incl. cross-connection consolidation
Feature flags	various	`SQUISH_DECAY_ENGINE`, `SQUISH_CRON_ENABLED`, `SQUISH_MULTIMODAL_ENABLED`

Architecture

Three-Layer Memory Model

Connectors

Squish connects to your existing tools and ingests context automatically:

Connector	What It Ingests
Google Drive	Documents, sheets, slides, and files
GitHub	Issues, PRs, discussions, code context, and repo metadata
Slack	Messages, threads, channel context, and decisions
Notion	Pages, databases, docs, and wikis

Connectors are available on Cloud tiers. Install with:

squish connect google-drive
squish connect github
squish connect slack
squish connect notion

Squish Cloud

Persistent memory across ChatGPT, Claude Desktop, Claude Code, and local agents. One account, synchronized everywhere.

Cloud features: OAuth 2.1 + PKCE login, cross-platform sync, team workspaces, admin dashboard, priority support.

Pricing

Tier	Price	Features
Local	Free forever	SQLite, 7 MCP tools, offline, knowledge graph, multimodal ingestion
Cloud Solo	$9/mo	Everything in Local + cloud sync, 1 connector, 10K requests/mo
Cloud Pro	$29/mo	Cross-tool sync, 3 connectors, 50K requests/mo, shared workspaces
Cloud Team	$99/mo	Unlimited seats, all connectors, 200K requests/mo, RBAC, audit logs

Sign up at squishplugin.dev -- 30 seconds, no credit card needed.

Benchmarks

Squish is tested against real-world memory retrieval tasks and synthetic benchmarks.

Metric	Result	Notes
Core Tests	9/9 passed (100%)	All memory operations
LoCoMo Memory	65%	100 REAL questions from locomo10.json
Throughput	39 ops/sec	With local embeddings
Total Time	230ms	For 9 core tests
Package Size	674 KB	Lightweight footprint
Latency (embed)	6.6ms	Local TF-IDF embeddings
Latency (search)	6.1ms	Hybrid retrieval

Full benchmark details: docs/BENCHMARK.md

Documentation

Document	Description
CLI Reference	All CLI commands and options
MCP Server	15 MCP tools and configuration
Architecture	System design and data flow
Decay System	How memories age and lose relevance
Scoring	Importance and relevance scoring
Environment Config	Environment variables and settings
Plugin Architecture	Hook system and agent integration
Quick Start	Getting started guide
Agent Comparison	Squish vs other memory tools
Contributing	How to contribute
Release Notes	Changelog and version history

FAQ

What is Squish?

Squish is memory infrastructure for AI agents. It provides a TypeScript SDK, an MCP server, and a CLI for persistent agent memory. Use the SDK to build memory-powered apps, plug the MCP server into any compatible agent, or use the CLI for quick operations. Local-first with optional Cloud sync. Think of it as a brain that persists between sessions -- your agents remember decisions, constraints, preferences, and context without you having to re-explain everything.

Does Squish require an API key?

No. Squish works locally by default with zero API keys. It uses local embeddings (TF-IDF) and SQLite storage. You can optionally configure an external LLM for enhanced reasoning, but it is not required. An API key is only needed if you want to use the paid Squish Cloud for cross-device sync.

How does Squish compare to mem0 or agentmemory?

Squish is the only option that works locally with zero external dependencies AND provides a TypeScript SDK. mem0 requires Qdrant (a vector database) and cloud API calls. agentmemory requires iii-engine. Letta and Zep require Postgres. Squish uses SQLite and local embeddings by default. See the full comparison in the Why Squish section above.

Can I use Squish with multiple AI agents?

Yes. Squish works with any MCP-compatible agent. One memory server is shared across Claude Code, Cursor, Codex, Copilot, Gemini CLI, and any other agent that supports MCP. Memories are available to all connected agents. You can also use the SDK directly in any TypeScript/JavaScript environment.

Is my data private with Squish?

Yes. In local mode, all data stays on your machine in an encrypted SQLite database. Nothing is sent to any cloud service. AES-256-GCM encryption protects sensitive memories. Cloud sync is opt-in and additive: data moves over TLS, connector credentials are encrypted at rest with AES-256-GCM, and synced memories live in the cloud database under your account's access controls -- which is exactly why local remains the default and sync stays optional.

What is the difference between recall and sessions?

squish recall searches your long-term memory -- distilled facts, decisions, and preferences that Squish has captured and organized. squish sessions search searches raw past agent runs -- the actual messages, commands, and file changes from previous Claude Code, Codex, OpenCode, or Gemini CLI sessions. Recall gives you what the system decided to remember. Sessions give you the evidence.

Every recall result carries a recall assessment with a verdict: confident (top result is reliable), qualified (usable but treat with care), or no_reliable_memory (nothing trustworthy found -- say so rather than guessing). Feed verdicts back via squish_feedback: confirm, used, or contradict.

Contributing

See docs/CONTRIBUTING.md for guidelines on how to contribute to Squish.

License

MIT -- see LICENSE for details.

Star the Repo

If Squish helps your project, consider starring the repo. It helps other developers find memory tools for their AI agents.

Website · Cloud Dashboard · Documentation · GitHub · npm

Name		Name	Last commit message	Last commit date
.github		.github
.husky		.husky
assets		assets
benchmark		benchmark
bin		bin
config		config
core		core
db		db
docs		docs
packages		packages
plugin		plugin
scripts		scripts
skills		skills
tests		tests
webui		webui
.dockerignore		.dockerignore
.env.example		.env.example
.gitignore		.gitignore
.npmignore		.npmignore
CHANGELOG.md		CHANGELOG.md
Dockerfile		Dockerfile
GITHUB_RELEASE_v1.6.0.md		GITHUB_RELEASE_v1.6.0.md
LICENSE		LICENSE
README.md		README.md
bun.lock		bun.lock
bunfig.toml		bunfig.toml
config.ts		config.ts
glama.json		glama.json
mcp.json.example		mcp.json.example
package.json		package.json
pnpm-workspace.yaml		pnpm-workspace.yaml
server.json		server.json
tsconfig.json		tsconfig.json

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

Squish - Memory Infrastructure for AI Agents

Get Started in 30 Seconds

Build on Squish

Pick Your Agent

Claude Code

Codex CLI (OpenAI)

Cursor / Windsurf / Cline

OpenCode

Any MCP Client

What Just Happened

Works with Every Agent

Why Squish

Comparison

What Makes Squish Different

Core Concepts

Features

Memory Intelligence

Multimodal Memory

Session Search

Interfaces

Storage

Self-Hosting

Environment Flags

Architecture

Three-Layer Memory Model

Connectors

Squish Cloud

Pricing

Benchmarks

Documentation

FAQ

What is Squish?

Does Squish require an API key?

How does Squish compare to mem0 or agentmemory?

Can I use Squish with multiple AI agents?

Is my data private with Squish?

What is the difference between recall and sessions?

Contributing

License

Star the Repo

About

Topics

Resources

Code of conduct

Contributing

Stars

Watchers

Forks

Releases

Sponsor this project

Packages

Contributors

Languages