Your AI’s memory, made visible.

A universal memory layer for AI agents. Searchable, correctable, portable. Install once, remember everywhere.

Get StartedView on GitHub
{
  "mcpServers": {
    "engrams": {
      "command": "npx",
      "args": ["-y", "engrams"]
    }
  }
}

Add to Claude Code, Cursor, Windsurf, or any MCP client. That’s it.

Your AI remembers. You can’t see what.

Invisible

Your AI builds up knowledge about you over time, but there's no way to see what it's stored. You'd have to ask, and trust the answer.

Fragmented

Teach something in one tool, start over in the next. Your context lives in separate walled gardens with no way to connect them.

Unaccountable

When your AI gets something wrong, you can't trace where it learned it or how sure it was. And corrections don't stick. You're left repeating yourself.

Engrams makes AI memory yours.

Search & Retrieve

Full-text search finds the right memory even when the wording differs. Filter by domain, confidence, entity type, or just ask.

› memory_search "deployment preferences"
Prefers blue-green deployments over rolling updates0.92
Uses GitHub Actions for CI/CD, not Jenkins0.87
Always runs smoke tests after deploy0.78

Correct & Control

Confirm what's right. Correct what's wrong. Split compound memories. Flag mistakes. Your AI learns from your feedback.

"Prefers TypeScript with strict mode enabled"

0.92
ConfirmCorrectSplit

Connect & Understand

Memories form a knowledge graph. People, projects, and preferences are automatically linked. Entity types extracted. Contradictions detected.

personprojectprefgoal

20 MCP tools. One install.

Everything your agent needs to remember, learn, and forget.

Core

memory_search

Full-text search with domain and confidence filters

memory_write

Create a memory with content, domain, and source

memory_update

Modify content, detail, or domain

memory_remove

Soft-delete with reason tracking

Trust

memory_confirm

Mark a memory as verified and boost its confidence

memory_correct

Replace content and reset confidence

memory_flag_mistake

Degrade confidence on bad information

Graph

memory_connect

Create typed relationship between memories

memory_get_connections

Traverse the knowledge graph

memory_split

Break compound memories into atomic units

Discovery

memory_list

Browse by domain, confidence, or recency

memory_list_domains

List all domains with memory counts

memory_list_entities

Show extracted entity types

memory_classify

Auto-classify a memory's entity type

Safety

memory_scrub

Detect and redact PII from memory content

memory_set_permissions

Per-agent read/write access by domain

Onboarding

memory_onboard

Personalized plan to seed your memory

memory_import

Import from Claude, ChatGPT, Cursor, gitconfig

memory_configure

Configure LLM provider for entity extraction

Pro

memory_sync

Sync memories with cloud across devices

Local-first. Zero config.

~/.engrams/engrams.db  ← Your memories live here. On your machine.

SQLite + FTS5 + sqlite-vec. No external database.

Embeddings run locally via Transformers.js. No API calls for search.

0600 file permissions. OS-level access control.

JSON export. Your data is always portable.

Optional cloud sync (Pro). AES-256-GCM encrypted, zero-knowledge.

Get started in 30 seconds.

~/.claude.json

{
  "mcpServers": {
    "engrams": {
      "command": "npx",
      "args": ["-y", "engrams"]
    }
  }
}

Same config for every client. Just change the file path.