Memorable

Teams running AI agents (coding, browser, voice, computer-use, ops) see those agents repeat the same work and re-derive the same solutions every session, wasting tool calls, tokens, and time. Today agents either have no memory or rely on vector search that returns loose chunks of text rather than a verified sequence of steps. Context windows fill up with large skill files (the page cites 15,593 tokens for an /investigate skill versus 293 for a learned procedure), and there is no record of which path actually worked or how it was verified. This costs extra tool calls, higher token spend, and lower task completion rates when agents repeat failed or inefficient paths.

Memorable watches an agent's own tool-call traces, strips out dead ends, and synthesizes successful runs into reusable 'procedures' with preconditions, postconditions, and a verify command. These procedures are assembled into a graph so shared steps across workflows reveal new paths and implicit skills, and a retrieval layer recalls the right procedure in about 60ms using exact, lexical, then semantic matching. It installs via a CLI (npx memorable-cli), works with Claude Code, Codex, Cursor, gbrain, QM, or any harness via a JSON trace to POST /v1/extract, and stores data locally, in a gbrain database, or in the user's own Postgres, sending only the prompt and allow-listed tool arguments off the machine. The dashboard shows every stored procedure along with time, token, and money savings per run, and procedures recorded by one agent can be recalled by any other agent sharing the same store.