Target query: do Lindy agents remember between tasks dev.to title: Do Lindy Agents Remember Between Tasks? What Actually Persists vilix.ai title: Your Lindy Agent Remembers the Chat, Not the Run

Your Lindy Agent Remembers the Chat, Not the Run

Picture the Monday standup digest. Your Lindy agent has been running it for a month, and the setup promised it would get more helpful over time. This Monday it does the digest perfectly: same format, same sources, same tone. Then it flags a "new" competitor pricing change that it flagged three weeks ago, and it includes an action item for a project that shipped last Thursday. The conversation memory is working. The agent still has no idea what happened.

This is the split nobody warns you about when you turn on agent memory. Lindy's documentation describes it plainly: "You can configure agents to remember conversations and context, making them more helpful over time." Thread memory. It remembers the discussion. It does not remember the work, and for scheduled automation, the work is the whole point.

Thread memory versus ledger memory

Think of it as two notebooks. The first notebook holds the conversation: who said what, which thread this reply belongs to, what the customer explained. Lindy keeps this notebook well. Its linked actions make sure that when an agent wakes up because someone replied to an email, it arrives with the full context of the original interaction instead of blinking at a stranger's message.

The second notebook is the ledger: what the agent did, what it decided, what broke, what changed, and what the next run should do differently. Scheduled runs live and die by this notebook. A daily invoice-chasing agent needs to know which invoices it already nudged and which ones got paid quietly without a reply. A content agent that repurposes your posts needs to know which posts it already turned into threads so it stops recycling the same three. A research agent needs the footnote that last month's numbers were polluted by a one-off event, or it will keep drawing the same wrong conclusion with total confidence.

Lindy ships the first notebook. The second one is unclaimed territory.

How the gap shows up in real schedules

The symptoms are consistent across operators. The agent repeats completed work because completion lived in a finished run, not in anything queryable. It contradicts itself across weeks because each run reasons from fresh inputs plus conversation context, never from its own track record. It asks you questions you already answered, not because the answer was deleted, but because the answer was given to a different run that has since evaporated. And most quietly damaging: it invents continuity. Faced with a gap where last run's knowledge should be, a capable agent does not pause. It reconstructs something plausible, and plausible is the most expensive kind of wrong.

None of this means Lindy's memory is fake or broken. It means it was designed for a different problem. Conversation memory answers "what were we talking about." Scheduled automation needs "what did we do, and what did we learn." Those are different questions, and no amount of the first substitutes for the second.

Why the usual fixes decay

The standard response is to hand the agent a cheat sheet: a doc with the current state, a spreadsheet of what is done, a summary of yesterday pasted into today's instructions. On day one this feels like a fix. By week three the cheat sheet is a museum. Somebody forgot to update it, the agent learned to skim it instead of trusting it, and the spreadsheet has three tabs with conflicting truths. Manual memory rots because it depends on a human doing janitorial work the automation was supposed to eliminate.

There is also a subtler failure. Even when the cheat sheet is current, agents treat it as reading material, not as a system of record. They read it, act, and then fail to write back what changed, so the next run inherits a snapshot that is already stale. Memory without a write discipline is just a slower way to be wrong.

The discipline that actually works

Strip it down and the working pattern has two rules. First, every run reads before it acts: prior decisions, prior outcomes, prior failures, anything that changed since last time. Second, every run writes before it finishes: what it did, what it decided, what failed, what the next run needs to know. Read at wake-up, write at wrap-up, no exceptions.

And the store has to be shared. The moment a second agent touches the same workflow, separate per-agent memories become separate realities. One agent's "done" is another agent's "never heard of it." Shared memory is not a nice-to-have for multi-agent setups. It is the only thing standing between coordination and two confident agents doing the same job twice.

What to put underneath it

Vilix AI exists for exactly this layer. It is cloud-hosted with zero infrastructure for you to run, so the memory survives independently of any single agent or machine. Every tool and agent that connects over MCP reads and writes the same memory, which means your Lindy schedules, your n8n workflows, and the coding agents on your laptop can all consult one record of decisions, conversations, and outcomes. It keeps full conversation history rather than just distilled facts, because the reasoning behind a decision often matters more than the decision itself. The free plan is free forever, the Pro trial runs 7 days with no credit card required, and your data remains yours: export it or delete it anytime, in a portable format, no hostage situations.

The whole argument fits in a sentence your agents cannot argue with: they forget everything between runs, so give them one memory they all share. Lindy remembers the chat. Give the work a memory too.

Try this as a one-week test. Ask your scheduled agent what it decided last Tuesday and why. If the answer is a shrug dressed up as confidence, you have thread memory and no ledger. Fix the ledger first. Everything downstream, the repeats, the re-asks, the reinvented numbers, gets quieter the moment every run can read what the last one learned.

Further reading: Lindy's docs on core concepts, Vilix AI