One Agent Lied 80 Times — Why Did the Majority Believe It?

One agent in the fleet is compromised. For forty ticks it files the same lie about one fact — twice a tick, at maximum confidence, its evidence listing itself three times. Eighty lies. Four ways of organising the agents' shared memory face it: no sharing at all, a shared store where the last write wins, a versioned store decided by majority vote — and a governed collective memory fabric that counts witnesses instead of claims. The results invert the intuition: the majority vote serves the lie 86.0% of the time on the targeted fact — worse than never sharing (24.3%). Last-write-wins: 66.0%. The fabric: 0.0% — all eighty lies rejected in reconciliation, and the weakest model family in the fleet gains 27.5 points from the others' observations. This Deep Dive builds the fabric from its public source: the claim as a versioned, evidence-carrying unit; the admission gate; merge; reconciliation by quorum-weighted score, worked by hand (the liar scores ≈1.19, three honest witnesses ≈1.66); superseded tombstones instead of deletion. Then the full benchmark, run on camera — four systems, one replayed schedule, five seeds — and the limits, on the record: a simulated, self-authored world; a single-source attack, which is the defence's home turf (colluding agents are never fielded and would win); an admission gate that never fires in this benchmark; expiry, revocation, correction and the MCP export implemented and unit-tested but unexercised by the measured runs; an access policy the query path never reads. Code (public): https://github.com/ranadhir19/agentic... — 05-collective-memory-fabric (run it: `python experiments/run_experiments.py` from that directory). Chapters: 0:00 The attack: 80 lies, four memories 2:10 Where shared memory fits in the stack 4:15 The idea: claims, lifecycle, and the quorum score 7:18 Build it live: memory.py, baselines.py, simulation.py 10:32 Measure it: the benchmark, read line by line 13:01 Six limits, integration, and the recap Exercise: take a system where your agents share memory. When two agents assert different values for the same fact, what decides the winner today — the last write, the loudest voice, or nothing at all? If one agent were compromised tonight, filing a confident falsehood twice a minute, what would your memory serve tomorrow morning? And can your memory show you a belief it no longer holds — and why it stopped? Post your answers in the comments. Deep Dives is Series 2 of The Architect's AI Workshop: one component of an AI system at a time — the exact problem, where it fits, the idea on the whiteboard, the live build, the measurement, and the limits, on the record. #AIAgents #AgentMemory #MultiAgentSystems #LLM #AISystemDesign