I Gave My Agents a Memory. Two Months of Real Data.

First, I stopped pasting

Before Echo, my actual job was carrying boxes between coworkers with amnesia. Research done in ChatGPT, carried into Codex. Architecture mapped in Claude Code, re-explained to the next session. A path ruled out yesterday, re-litigated today.

Each dot ≈ 1,700 characters of context per session. I used to paste all 212. Now I type the blue one — my agents retrieve the rest through Echo.

÷212

context I no longer carry by hand

360,000→1,700characters per session

My worst single session before Echo: 798,000 characters pasted.

ChatGPT

Claude Code

Codex

One memory layer under every agent.

My hand-carried context fell from ~360,000 characters per session to ~1,700. The agents’ own retrievals rose from ~2 to ~15 per turn. The job of moving the past into the present left my hands.

I am no longer the memory bus between my AI tools.

The agent got smarter — starting with its first move

Without memory, the first half of every fresh session is archaeology: git log, rg, re-reading the same files, waiting for me to correct its understanding. With memory, it starts from the decisions, constraints, and dead ends that already exist — and the change shows up in the logs, not in adjectives.

This is not memory making the model smarter. It is the model no longer walking through yesterday’s ruins to reach today’s problem.

Not a smaller bill. A bigger day.

If the story ended at “tokens down 63%,” Echo would be a cost optimization. The real question is where the saved hauling went.

And the number nobody expects: my research and GTM sessions now burn more tokens and more tool calls than before. That is not regression — that work used to be done by me, by hand, at midnight. Now agents actually do it.

I do not work longer hours. My work just stopped resetting to zero every morning.

The harness thesis

Models are standardizing. What decides your results is the harness around them — the context, memory, tools, permissions, and evals the model works inside. We unpacked the mechanics in How Agents Work: A Deep Dive Into the Context Window. Inside that harness, memory decides whether an agent works continuously or resets to zero every morning — and it is the hardest layer to retrofit later.

It’s not a smarter model. It’s a harness that stops throwing away what the model already learned.

Source:Codex828 sessionsClaude Codefull transcripts · my own production work, not a benchmark

Method: the watershed is June 12, 2026 — my first real EchoMem MCP call. Data covers 828 Codex sessions plus local Claude Code / Desktop transcripts. Coding numbers use same-repo paired history (5 Echo warm-up sessions vs 7 no-Echo controls, per user turn), not a randomized same-model A/B. Non-coding numbers compare a workflow migration: chat + manual pasting became agent + memory recall. All of it is my own production work, not a general benchmark. 中文版：我给 Agent 接上了记忆。这是两个月的真实数据。