Base-native AI memory infrastructure
FlowMemory combines Rootflow and FlowCodec: Uniswap v4 swaps create verifiable pulses, then compressors turn accepted work into proof-carrying memory capsules for AI agents.
The category error
The hook stays boring: no custody, no custom accounting, no fee mutation, no swap delta changes. The intelligence happens after the pulse, in the verifiable work and memory layers.
normal swap
-> afterSwap pulse
-> checkpointed cursor
-> worker artifact
-> verifier quorum
-> storage proof
-> memory capsuleTwo engines, one product
Rootflow turns swap pulses into checkpoint-backed cursor proofs, worker jobs, subject-bound verifier quorums, adapter-backed storage proofs, canonical roots, and reserved payouts.
FlowCodec turns epochs into TracePointers, epoch roots, compressor contests, validity scores, and proof-carrying MemoryCapsules that AI agents can query and reconstruct.
Proof of Forgetting
FlowCodec does not save every log as text. It saves compressed state, surprise residuals, regime sketches, proof roots, and reconstruction instructions. Exact recall comes from chain evidence only when needed.
Build status
Base mainnet canary completed with real v4 pool, FlowPulse, checkpoint, cursor-proof job, quorum, settlement, and reserved reward payout.
Local v4-core demo, indexer, TracePointers, MemoryCapsules, query API, and Memory AMM scaffolds are running. Public Base deployment is next.
Production storage adapters, verifier governance, source verification, Safe custody, audit, and launch approval remain before public mainnet.
FlowMemory