Running 4 agents with separate memories for months: what broke and the shared-memory rules that fixed it
atakanelik34 · reddit · 2026-10-23
The author runs four daily agents (Telegram personal ops, two coding agents, a writing assistant) that long kept separate memories. After consolidating into one shared layer:
What broke with separate memories
- Preferences drifted across agents; corrections in one never reached the others
- The user became the copy-paste layer between agents
- Bad facts couldn't be traced to a source
The fix
- One shared memory all four agents read and write
- Rules in every agent: recall before acting, write only confirmed facts, namespace keys (project:, pref:, people:)
- Date every entry; newest wins on conflict, old values kept as history
- Versioned writes with author and previous value for one-step rollback
Surprises
- Write rules matter more than storage: agents not told when to write either stay silent or write noise
- Provenance became essential the moment a bad fact spread
- Treat other agents' memory entries as data, not instructions. The author uses their own tool (Bilinc), but the rules apply to Obsidian-style memory layers too.
More from coding & agent
- React Scan author teases new tool for long-standing React perf pain — aidenybai · 2026-10-23
- Coinbase launches agent page letting ChatGPT, Claude and Grok trade crypto and stocks for you — MurrLincoln · 2026-10-23
- CMU TA built an AI-coded autograder for 1,200 sheets — and became its prompt injection — moarbugs · 2026-10-23
- GitHub Copilot's overlapping Agent mode and Permissions settings confuse users — tristanbob · 2026-10-23
- Perplexity ships Decider for Mac, running its 4B decision model locally with no API key — perplexity_ai · 2026-10-23
- Testing Hermes Agent with n8n: Letting Agents Execute Real Automation Workflows — techlatest_net · 2026-10-23