How I Gave My AI Coding Agent Persistent Memory in 2 Minutes

#ai #devtools #mcp #productivity

How I Gave My AI Coding Agent Persistent Memory in 2 Minutes

Every AI coding session starts the same way: you explain your project, your conventions, your preferences, and the context from last week's debugging session. Then you do it all again tomorrow.

I built Engram to fix this. It gives ChatGPT, Claude Code, Cursor, and Windsurf a persistent memory that survives across sessions — so your AI actually remembers what you discussed yesterday.

The Problem

AI assistants are stateless. Each session starts from zero. There are workarounds:

CLAUDE.md / .cursorrules — manual, limited, you maintain it by hand
ChatGPT's built-in memory — stores brief facts like "user likes Python," not the full reasoning behind decisions
Copy-pasting context — tedious, doesn't scale

None of these capture the full arc of a conversation — the trade-offs you evaluated, the bugs you investigated, the reasoning behind decisions.

What Engram Does

Engram stores your conversations verbatim and makes them searchable by meaning. When your AI starts a new session, it can search your history and pull in relevant context from days or weeks ago.

It works through MCP (Model Context Protocol), so any compatible client can use it. Your AI gets three core tools:

search — "what did we decide about the auth flow?" finds the relevant conversation even if you never used those exact words
create_conversation / append_messages — stores new conversations automatically
vault_set / vault_get — key-value store for structured facts

Under the hood, conversations are chunked and embedded using vector embeddings. Search is semantic — it matches by meaning, not keywords.

Setup: ChatGPT (30 seconds, no install)

Go to Settings > Apps > Developer mode
Add a custom MCP connector: https://mcp.getengram.app/mcp
Sign in with OAuth when prompted

That's it. No CLI, no API keys, no install. ChatGPT can now search your history and store new conversations.

Try asking: "What did we discuss about [topic] last week?"

Setup: Claude Code (one config line + auto-capture)

Add to your MCP config (~/.claude/settings.json):

{
  "mcpServers": {
    "engram": {
      "type": "http",
      "url": "https://mcp.getengram.app/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}

Then add a few lines to your CLAUDE.md telling Claude to search on session start:

## Memory

On session start, search Engram for context relevant to the current task:
- search query: "<summary of what the user is asking about>"

When important decisions or context are established, store them in Engram.

Auto-capture with the CLI daemon

For zero-friction capture, install the CLI:

npm i -g @getengram/cli
engram daemon start

The daemon runs in the background and auto-captures every Claude Code session. You don't have to prompt Claude to remember things — it just happens.

Setup: Cursor / Windsurf

Same as Claude — add the MCP server URL + API key to your MCP config. Both support MCP natively.

What It Actually Remembers

After a few weeks of use, here's the kind of context Engram surfaces for me:

"We set up OAuth 2.1 with Dynamic Client Registration for the ChatGPT integration"
"The rate limiter is 100 req/min per org, configured in the Worker binding"
"User prefers explicit error handling, no swallowed exceptions"
Past bug investigations so I don't chase the same issue twice
Architecture decisions and the full reasoning behind trade-offs

The key difference from CLAUDE.md: I don't maintain any of this manually. Conversations are stored as they happen, and the right context surfaces when it's relevant.

How It Works (for the curious)

The backend runs entirely on Cloudflare:

Workers — API + MCP server (Hono.js)
D1 — SQLite storage for conversations
Vectorize — vector index for semantic search
Workers AI — embedding generation (bge-base-en-v1.5)

When you store a conversation, Engram:

Splits it into chunks (~500 tokens each)
Generates vector embeddings for each chunk
Indexes them in Vectorize

When you search, it:

Embeds your query
Finds the most similar chunks via cosine similarity
Returns the relevant conversation snippets with full context

The whole thing is globally distributed with sub-100ms cold starts.

Pricing

Free — 1,000 messages/month (enough to try it out)
Pro — $39/month for heavier use
Team — $49/seat/month with shared memory across your team

Source Code

The source is on GitHub under BSL-1.1 (converts to Apache 2.0 after the change date). You can self-host the whole thing on Cloudflare's free tier with wrangler deploy.

GitHub: github.com/get-engram/engram