# I gave up on perfect agent memory and everything got cheaper — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/e40977f0ff36381d

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-09-27 13:41 ingested 2026-09-28 02:36

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1wrkja1/i_gave_up_on_perfect_agent_memory_and_everything/)

## retained text



```
For a while I was doing the full capture thing. Every session got summarized and injected into the next one, the whole history, nothing lost. Felt like the right approach, like how you'd keep notes after a meeting.

Two problems. First it was expensive, I was burning like 10k+ tokens before the agent wrote a single line, just re-loading its own diary. Second, and this was worse, the summaries would confidently carry forward stuff that was wrong. A wrong decision from three weeks ago would show up as settled fact and the agent would build on it.

What I do now is dumber. One short file, I edit it by hand, only things I'd actually tell a new hire. Preferences, gotchas, the two or three architectural decisions that matter. Maybe 40 lines. The agent reads it at session start and that's it.

Token bill dropped a lot. And weirdly the agent got more reliable, because now when it doesn't know something it asks instead of trusting a stale summary.

Anyone else find that less memory works better? Or did I just implement the capture thing badly?
```



document fetched 2026-09-28 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1wrkja1/i_gave_up_on_perfect_agent_memory_and_everything
