# I cut my token spend 50%+ by adding a shared memory layer across agents — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/68f9f5818784f046

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-09-30 12:00 ingested 2026-10-01 02:37

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1wu31ir/i_cut_my_token_spend_50_by_adding_a_shared_memory/)

## retained text



```
I run an orchestrator setup where one model plans and delegates and the workers execute, before adding memory every single call to the orchestrator was replaying the full project state, what's been done, what's pending and more re injected every turn and over 60% of my orchestrator's tokens were just re reading context it already had from the last call and my bill was give or take $580/month before this.

The workers were worse bc they had zero visibility into what other workers did so agent 3 would rewrite something agent 1 already built and the orchestrator would re delegate tasks it already sent out 20 calls ago because it lost track.

now I didn't have many options atp, I tried adding shared memory layer so all agents read from and write to the same store scoped per project. The orchestrator gets a ranked summary of current state instead of the full history every call and workers check what already exists before they build anything and token spend dropped by half and the duplicate work basically stopped.

my bill now is about $260/month and the memory layer costs me $19/month so net I'm saving $300/month which is good enough for me.

If you're running multi agent and your bill feels high check how much of each call is just context re injection bc that was my whole problem.
```



document fetched 2026-10-01 00:20 · canonical https://reddit.com/r/AI_Agents/comments/1wu31ir/i_cut_my_token_spend_50_by_adding_a_shared_memory
