roampal.ai

What It Is

Roampal is a persistent memory layer for AI coding tools, focused on solving the problem where assistants such as Claude Code and OpenCode “forget” past decisions, user preferences, and fixes that actually worked. It does not replace the underlying large language model; instead, it maintains a local memory store and automatically injects, records, and scores context before and after each conversation turn.

Core Capabilities

Its technical approach is more sophisticated than a basic similarity-based RAG setup: for each exchange, it generates a summary, extracts atomic facts, and derives noun-based tags. At query time, TagCascade first narrows candidates by tags, then reranks them with a cross-encoder. By default, each turn injects 4 summaries and 4 facts, along with metadata such as used, last, and wilson. Memories are scored based on outcomes such as “worked / failed / partial” and are promoted, demoted, or expired across layers such as working, history, and patterns. According to official LoCoMo test results, TagCascade achieved 76.6% accuracy, improving by 23.6 points over raw ingestion at 53.0%; in memory-poisoning tests, performance dropped by 2.6–4.2 points. These are official figures, however, and real-world performance will still depend on the model and usage scenario.

Pricing and Privacy

roampal-core is free and open source under Apache 2.0. Desktop is sold through Gumroad, but the text does not disclose a specific price. Privacy is one of its strengths: ChromaDB stores data locally, and result records are written to local metadata. Desktop can use Ollama or LM Studio for local inference. The official claim is zero telemetry: data does not leave the machine, with only minimal network access required for things like checking the PyPI version at startup and downloading the embedding model on first use.

Integrations, Pros, and Cons

Core offers the best experience with Claude Code and OpenCode, where hooks/plugins can automatically inject context and enforce scoring. Desktop supports MCP-compatible tools, can auto-detect configurations, and provides tools for search, add, update, and scoring. However, it lacks Core’s hooks-based automatic injection and mandatory scoring, so it relies more on users prompting it manually. Its strengths are that it is local, open source, transparent in its mechanisms, and well suited to long-running projects. Its weaknesses are that it is geared toward developer toolchains, has a relatively high setup and MCP learning curve, lacks transparent Desktop pricing, and offers limited value outside programming scenarios.

Who It’s For and Access from China

Roampal is best suited to heavy Claude Code/OpenCode users, developers who want AI to remember project conventions, and individuals or small teams that care about keeping data local. Access from China is not specified in the text. If it depends on Gumroad, PyPI, GitHub, or overseas model services, network connectivity and payment may be uncertain. Alternatives include the built-in context features of Cursor/Windsurf, Continue, Pieces, or building a custom ChromaDB/RAG memory layer.