Recoverable shared memory for multi-agent setups: files, append-only logs, provenance
RocketSeven · reddit · 2026-09-30
A Reddit poster asks how to make shared agent memory recoverable after a bad write, noting that neither Markdown alone nor a vector index solves the case where one agent writes a confident but wrong summary that later runs build on top of.
Their leaning design:
- Human-readable files as source of truth, immutable snapshots or append-only decision log, and a derived index that can always be rebuilt
- Every memory update carries author, timestamp, source links, status, and a supersedes field
- Concurrent writes fail on stale revisions instead of silent last-write-wins
Open questions: how to distinguish facts, decisions, temporary observations and generated summaries in practice, and what minimum design lets a human answer what agents believed at a given time, why, and how to restore the last trustworthy state.
More from coding & agent
- BRIDGE: bilevel RL jointly trains search agent LLM and retriever, boosting multi-hop QA by 9.6 EM — _reachsumit · 2026-09-30
- Anthropic engineer: stop prompting Claude, build a system that prompts itself — s_mohinii · 2026-09-30
- CorpusMap paper builds offline entity pages so agents stop re-searching the corpus — _reachsumit · 2026-09-30
- Burkov Reorders ChapterPal's Model Fallback Chain, Crowns GPT-6 Luna Best Value — burkov · 2026-09-30
- The log you can trust after a week unattended is the one the agent never wrote — jkris050 · 2026-09-30
- Cube Launches Always-On Cloud Computers for Running Claude Code and Codex Agents — algo_diver · 2026-09-30