Long-term memory
for Cline.
Every Cline session becomes permanent. Your memory bank updates itself. Your team’s patterns become skills. Your data trains your model. One capture. Four rungs.
The Problem
Cline is stateless between sessions. The official workaround is Memory Bank: six markdown files — projectbrief.md, productContext.md, activeContext.md, systemPatterns.md, techContext.md, progress.md — that the user maintains by hand and refreshes with “update memory bank” before every context reset. Worse, every session expires soon after, so the trainable record disappears with it.
Paper Compute
Automates what Memory Bank asks you to do by hand — and turns the same captures into team skills and training data.
Captures every Cline session at the network layer and keeps your memory-bank/ in sync without prompts. No SDKs. No changes to Cline itself. One env var.
Team-wide memory bank rollup, skill clustering, and the fine-tune pipeline. Every successful session accrues into a shared library and, eventually, your own weights.
One capture. Four rungs.
tapes sits between Cline and its model provider. Every session branches into four rungs: raw capture, an auto-maintained memory bank, a clustered skill library, and eventually a fine-tuned model. One env var for free users. One toggle in Cline Enterprise.
tapes Proxy sits between Cline and its model provider. Captures every LLM call, tool invocation, and file edit into a Merkle DAG. Ships as a single binary so every Cline surface picks it up the same way. No SDK. No code changes inside Cline.
Memory Bank Sync writes diffs back to your memory-bank/ folder after every session. Follows Cline’s Memory Bank spec. The user never has to type “update memory bank” again.
Clusters successful sessions across the team, extracts the common execution pattern, and drafts a SKILL.md file for Cline’s native skills system. Developers review and publish. Every team gets the accumulated knowledge of every win.
Accumulated sessions feed a fine-tune pipeline hosted in Paper Compute. The team gets a Cline-specific adapter trained on the way they actually ship — priced by compute, not by seat, and ready whenever the session volume crosses the SFT threshold.
The Ladder
Session data is the input. Each rung is a tier upgrade: raw capture, then an auto-updated memory bank, then a team skill library, then a fine-tuned model. Below is what each rung looks like in the Paper Cloud UI.
Partnership Rollout
A staged path from free integration to enterprise tier revenue. Each phase ships independently.
A live Cline session that writes back to memory-bank/activeContext.md and progress.md with no “update memory bank” prompt. Sessions stored in a Merkle DAG, fully replayable. Runs end-to-end in under 60 seconds.
Cline ships Memory Bank auto-sync as a single toggle that works everywhere Cline runs — the extension, the CLI, and Cline Kanban. Every user stops maintaining six markdown files by hand. Rungs 1 and 2 live.
Team-wide memory bank rollup, skill clustering, and compliance audit trails. Paper Compute prices by compute — storage, sync, and skill clustering scale with how much the team actually runs, not with seat count. Cline sells it as part of their enterprise pitch. Rung 3 live.
When a customer’s session data crosses the SFT threshold, Cline hands them to Paper Compute’s training pipeline — the same play Cursor ran with Composer 2, without Cline needing to own the editor. Cline gets a referral fee. Rung 4 live.
Why Now
Every coding tool loses one thing at the end of a session: what the agent actually figured out. The commits survive. The reasoning doesn’t. Cline’s Memory Bank is the first serious attempt to fix this — and it works, until someone forgets to type “update memory bank”.
tapes turns that workflow into infrastructure. Every session you run compounds into knowledge your team keeps: a memory bank that stays current on its own, a skill library the team can reuse, and eventually a model tuned on the way your team actually works. Nothing gets locked to a vendor — the capture is local, the format is open, and the data is yours.
This matters more with every surface Cline adds. Cline Kanban moves the agent out of a single editor pane and onto a board where multiple agents ship work in parallel. More agents mean more sessions, more reasoning worth keeping, and more patterns worth reusing. tapes is the layer that turns that volume into value instead of noise. Every Cline user ends up asking two questions:
Cursor showed that session data, once you keep it, becomes the raw material for real model improvement. tapes gives Cline users the same ingredient — a durable record of how their team actually ships. Every session a team runs today is still working for them a year out.
What Cline unlocks on paper.
tapes follows Cline’s Memory Bank spec and is AGPL-licensed. Session data is captured locally, stored in a content-addressed Merkle DAG, and stays in formats you can read, export, or migrate. The goal is data that stays useful to you over time — not data held in escrow by a vendor.