PRO-LONG

Durable memory for coding agents.

Long tasks outlive context windows. After compaction or a fresh session, coding agents can lose earlier decisions, repeat failed work, and miss important tool results.

PRO-LONG records coding-session events in one local, append-only log and gives the agent a skill for searching only the history it needs. You keep using your normal coding CLI—there is no wrapper, server, database, or transcript dumped into the prompt.

Works with Codex, Claude Code, OpenCode, and pi.

• August 19, 2026 — Added PRO-LONG memory for coding CLIs, with project-local integrations for Codex, Claude Code, OpenCode, and pi.

Install PRO-LONG, then initialize it inside your project:

git clone https://github.com/alexisfox7/PRO-LONG.git && cd PRO-LONG
npm install && npm run build && npm link

cd /path/to/your/project
prolong init

Then use your coding agent normally. PRO-LONG records in the background and the agent retrieves relevant history when a task spans compaction or sessions.

flowchart LR
    C["Coding agent"] -->|"lifecycle events"| L[".prolong/log.jsonl"]
    L -->|"search when needed"| M["Relevant prior context"]
    M --> C

1. prolong init installs a small project-local adapter and the prolong agent skill.
2. The adapter appends prompts, tool activity, assistant handoffs, and session boundaries to .prolong/log.jsonl.
3. When prior work matters, the skill teaches the agent to search the log with ordinary tools such as rg, jq, or Python and read only the relevant entries.

PRO-LONG does not inject the accumulated log into every prompt. Reads of the log are also excluded from recording, so retrieval cannot recursively copy the memory back into itself.

The complete CLI has three commands:

prolong init
prolong status
prolong uninstall

• init detects supported coding clients on PATH.
• status checks the skill, runtime, and client integrations. Use status --json for automation.
• uninstall removes the integration but preserves the log. Add --purge to delete the recorded history too.

To select clients explicitly:

prolong init --client codex,claude-code,opencode,pi

Hooks and extensions run with your user permissions. Review the generated integration and accept the client's project or hook trust prompt before use.

The log stays inside the project and is gitignored by default. It can contain prompts, assistant messages, tool inputs, tool results, and any secrets that appeared in them. Do not enable transcript retention where local policy forbids it. Run prolong uninstall --purge when the history should be removed.

PRO-LONG began as research on programmatic memory for long-horizon agents. On the full ARC-AGI-3 public game set, the research harness improved over matched coding-agent baselines by 18 percentage points on average and reached 97.4% best@2 with Fable 5. These results motivate the product design; they are not yet a direct benchmark of the coding-tool MVP.

The original PRO-LONG context framework evaluated on ARC-AGI-3.

• Paper: PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning
• ARC-AGI-3 reproduction code and logs
• Citation metadata

MIT