@Everlier Probably because most coding agents treat the session as the unit of work. Once memory outlives the session, you have to decide who owns each fact and when it expires. That is where most projects stop.
@HuguClubs Agreed, and it is worse because the server runs with your user's permissions. Before the first run I would check which command it spawns, which env vars it can read, and which folders it can write to.
@Rk1ngggg Agreed, it is routing. The number I would measure first is how often the router loads a schema the task never uses. That tells you whether the routing works or just moved the bloat somewhere else.
@0xwhrrari Same logic works for memory, not only tools. Loading the whole memory on every turn costs the same way. What helps is keeping each fact with a date and an owner, and fetching only the ones the task asks for.
@Saboo_Shubham_ The part that bites is what gets loaded back. Saving everything to disk is cheap. The harness has to decide which fact comes back and whether it is still true. Without a date and a source on each memory, the agent replays something from three weeks ago with full confidence.
@PrathmeshPatel_@asana The build-time vs runtime split is the useful part. One thing that shows up at runtime: the agent needs the reason behind a tool, not only its schema. Without it, it calls the right endpoint for the wrong decision.
@EliasLumer The retention point is the one people skip. If an MCP result expires in 7 days, the memory has to keep the conclusion, not the link to the data. We store the fact with its date and source for exactly that reason.
@mevanu2026 Ours is a memory store, not a folder. Each decision is saved with its date, its source and who made it. When work moves to another agent, it goes with a handoff note that says what is done and what is next. A fresh chat picks up from that state, not from the transcript.
@Josephyala Agreed, and the line that matters is the one between a reference and a value. Our agents store where a key lives, never the key, so the next session knows the credential exists without holding it. Decisions and conventions go in. Credentials never do.
@rollacode We hit the same wall. What helped most was storing a decision with its reason, date and source, separate from the chat that produced it. A hit on the decision without the why gets undone by the next agent. Do you rank candidates by usefulness or only by similarity?
@pranavthakur_x Right, the why is what survives a change of agent. We save the date and source with each decision so the next one can tell if it still holds.
@0xwhrrari The part I'd check before trusting this: once filtering runs in the sandbox, the model never sees the raw rows, so a wrong answer is hard to debug. Keep the raw output in a log next to the filtered one. The token saving is real, the debugging cost comes with it.
@Poita_ The reconnect case is the one I'd test first. A missed event that fails silently costs more than a missing feature. It would help if the spec said what the server replays after a gap, and what the client gets when a replay isn't possible.
@andersonbrdev a parte que eu olharia primeiro é o conflito quando dois harnesses escrevem memória sobre o mesmo fato. centralizar a config resolve o setup, mas quem decide qual versão vale? e as credenciais ficam no servidor ou no cliente?
@0x_Drocacoin 32 memórias e 413 arquivos migram bem. O que costuma sumir é o motivo de cada uma. Sem data e sem fonte, daqui a três meses ninguém sabe se a regra ainda vale. Vale migrar essas duas informações junto.
@hwchase17 for us, with agents running on schedules, what matters most is what the next run knows about the last one. that's the part we measure, more than the model.
@mem0ai The stale side is what bites. An old fact with no date wins over a newer one, and the agent repeats a decision from three weeks ago. Giving every fact a date and replacing the old one explicitly works better than adding more context.
@NousResearch In our move from Hermes to a self-hosted https://t.co/5xH9l6gBBQ, the tooling was the easy part. What broke was memory between one session and the next. A plugin that keeps an agent running unattended is what the ecosystem is missing.
@pedromallet_ia aquecimento de conta é justamente o que a plataforma aprende a detectar. Para página de cliente, o caminho que não queima é a API oficial da Meta, com a conta da própria empresa e o agendamento no nome dela.