Agents that don’t forget
By default, an AI session starts from nothing and ends the same way — every lesson learned dies with the process. For long-running automation, that’s expensive: humans re-teach the same recipe on every run.
Pinard gives its agents persistent memory. Through the mem_* tools (backed by Engram), an agent saves durable observations — a bug fix, an architecture decision, a config gotcha — and recalls them in later sessions. Knowledge accumulates instead of evaporating.
What gets remembered
- Decisions & patterns — why something was done a certain way, conventions to follow
- Fixes & gotchas — the recovery recipe for a failure the agent hit before
- Project context — durable facts about a repo, scoped per project so nothing bleeds across estates
mem_save(title="Deploy model: one shared checkout serves all vignobles",
type="config", scope="project")
mem_search("how do we deploy bin/pinard to vignobles?")
Replicated, not trapped on one host
Memory is local-first but replicates to a central store — so an agent on your workstation, a remote node, and an HPC job share the same accumulated knowledge. A background flush keeps the cloud copy current; nothing is stranded on the machine that happened to learn it.
On the horizon — a knowledge graph
The next vintage is a temporal knowledge graph: operational facts extracted from agent conversations, with validity windows so stale knowledge expires, and boot-time injection so a fresh worker starts already knowing its pipeline’s recipes. Teach once; the estate remembers.
The cellar’s memory
A great cellar is a library of vintages — every year’s triumphs and mistakes recorded, so the next harvest builds on the last. Pinard’s memory is that cellar book: the estate gets wiser with every season.